close

.github/workflows/build-single-tinygrad-model.yaml

@@ -12,11 +12,11 @@ on:
         required: false
         type: string
       recompiled_dir:
-        description: 'Existing recompiled directory number (e.g. 1 for recompiled1)'
+        description: 'Existing recompiled directory number (e.g. 3 for recompiled3)'
         required: true
         type: string
       json_version:
-        description: 'driving_models version number to update (e.g. 18 for driving_models_v18.json)'
+        description: 'driving_models version number to update (e.g. 5 for driving_models_v5.json)'
         required: true
         type: string
       artifact_suffix:
@@ -63,11 +63,12 @@ on:
         default: 'None'
         options:
           - None
-          - Master Models
+          - Simple Plan Models
-          - Release Models
+          - Space Lab Models
-          - 2025 World Models
+          - TR Models
-          - 2026 World Models
+          - DTR Models
           - Custom Merge Models
+          - FOF series models
           - Other
       custom_model_folder:
         description: 'Custom model folder name (if "Other" selected)'

.github/workflows/sunnypilot-build-model.yaml

@@ -30,11 +30,6 @@ on:
         required: false
         type: string
         default: ''
-      target_hardware:
-        description: 'Hardware target to compile for (qcom or usbgpu)'
-        required: false
-        type: string
-        default: 'qcom'
   workflow_dispatch:
     inputs:
       upstream_branch:
@@ -51,14 +46,6 @@ on:
         required: false
         type: boolean
         default: true
-      target_hardware:
-        description: 'Hardware target to compile for'
-        required: true
-        type: choice
-        options:
-          - qcom
-          - usbgpu
-        default: 'qcom'
 
 
 run-name: Build model [${{ inputs.custom_name || inputs.upstream_branch }}] from ref [${{ inputs.upstream_branch }}]
@@ -174,7 +161,7 @@ jobs:
           name: models-${{ env.REF }}${{ inputs.artifact_suffix }}
           path: ${{ env.MODELS_DIR }}
       - run: |
-          rm -f ${{ env.MODELS_DIR }}/{dmonitoring_model,big_driving_policy,big_driving_vision,big_driving_supercombo}.onnx
+          rm -f ${{ env.MODELS_DIR }}/{dmonitoring_model,big_driving_policy,big_driving_vision}.onnx
 
       - name: Build Model
         run: |
@@ -186,17 +173,7 @@ jobs:
           COMPILE_MODELD="${{ github.workspace }}/openpilot/sunnypilot/modeld_v2/compile_modeld.py"
           MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
           CAMERA_RES=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
-
+          TG_FLAGS="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
-          if [ "${{ inputs.target_hardware }}" == "usbgpu" ]; then
-            echo "USBGPU build"
-            export USBGPU=1
-            TG_FLAGS="DEV=AMD USBGPU=1 IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
-            OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
-          else
-            echo "QCOM build"
-            TG_FLAGS="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
-            OUTPUT_PKL="${{ env.MODELS_DIR }}/driving_tinygrad.pkl"
-          fi
 
           # Generate metadata for all ONNX files
           find "${{ env.MODELS_DIR }}" -maxdepth 1 -name '*.onnx' | while IFS= read -r onnx_file; do
@@ -209,13 +186,7 @@ jobs:
           POLICY_ONNX="${{ env.MODELS_DIR }}/driving_policy.onnx"
           OFF_POLICY_ONNX="${{ env.MODELS_DIR }}/driving_off_policy.onnx"
           ON_POLICY_ONNX="${{ env.MODELS_DIR }}/driving_on_policy.onnx"
-          SUPERCOMBO_ONNX=""
+          SUPERCOMBO_ONNX="${{ env.MODELS_DIR }}/supercombo.onnx"
-          for f in "${{ env.MODELS_DIR }}/supercombo.onnx" "${{ env.MODELS_DIR }}/driving_supercombo.onnx"; do
-            if [ -f "$f" ]; then
-              SUPERCOMBO_ONNX="$f"
-              break
-            fi
-          done
 
           MODEL_TYPE="" ONNX_ARGS="" OUTPUT_NAME=""
           if [ -f "$VISION_ONNX" ]; then
@@ -236,15 +207,24 @@ jobs:
           fi
 
           if [ -n "$MODEL_TYPE" ]; then
-            echo "Detected: $MODEL_TYPE -> $OUTPUT_PKL"
+            echo "Detected: $MODEL_TYPE -> driving_tinygrad.pkl"
             env ${TG_FLAGS} python3 "$COMPILE_MODELD" \
               --model-type $MODEL_TYPE \
               --model-size $MODEL_SIZE \
               --camera-resolutions $CAMERA_RES \
               $ONNX_ARGS \
-              --output "$OUTPUT_PKL"
+              --output "${{ env.MODELS_DIR }}/driving_tinygrad.pkl"
           fi
 
+      - name: Validate Model Outputs
+        run: |
+          source /etc/profile
+          export UV_PROJECT_ENVIRONMENT=${HOME}/venv
+          export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
+          python3 "${{ github.workspace }}/release/ci/model_generator.py" \
+            --validate-only \
+            --model-dir "${{ env.MODELS_DIR }}"
+
       - name: Prepare Output
         run: |
           sudo rm -rf ${{ env.OUTPUT_DIR }}

openpilot/cereal/custom.capnp

@@ -137,16 +137,10 @@ struct ModelManagerSP @0xaedffd8f31e7b55d {
     eta @2 :UInt32;
   }
 
-  struct Chunk {
-    fileName @0 :Text;
-    sha256 @1 :Text;
-  }
-
   struct Artifact {
     fileName @0 :Text;
     downloadUri @1 :DownloadUri;
     downloadProgress @2 :DownloadProgress;
-    chunks @3 :List(Chunk);
   }
 
   struct Model {
@@ -161,7 +155,6 @@ struct ModelManagerSP @0xaedffd8f31e7b55d {
       policy @3;
       offPolicy @4;
       onPolicy @5;
-      chunked @6;
     }
   }
 

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
+      nv12 = self.frame_buf_params[key]
-      yuv_size = self.frame_buf_params[key][3]
+      yuv_size = nv12[3]
-      # There is a ringbuffer of imgs, just cache tensors pointing to all of them
+      if self.copy_vision_buffers:
-      cache_key = (key, ptr)
+        # VisionIPC supplies a CPU pointer. Metal's external_ptr expects an MTLBuffer object,
-      if cache_key not in self._blob_cache:
+        # so wrap its NV12 pixels in the Venus layout expected by the compiled warp, then copy.
-        self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8', device=self.WARP_DEV)
+        if key not in self._vision_staging:
-      self.full_frames[key] = self._blob_cache[cache_key]
+          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"])
+  output_services = ["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"]
-  sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"])
+  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():
+      # if not self.sm.all_alive():
-        self.events.add(EventName.commIssue)
+      #   self.events.add(EventName.commIssue)
-      elif not self.sm.all_freq_ok():
+      # elif not self.sm.all_freq_ok():
-        self.events.add(EventName.commIssueAvgFreq)
+      #   self.events.add(EventName.commIssueAvgFreq)
-      else:
+      # else:
-        self.events.add(EventName.commIssue)
+      #   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.CP.notCar:
-      if not self.sm['livePose'].posenetOK:
+    #   if not self.sm['livePose'].posenetOK:
-        self.events.add(EventName.posenetInvalid)
+    #     self.events.add(EventName.posenetInvalid)
-      if not self.sm['livePose'].inputsOK:
+    #   if not self.sm['livePose'].inputsOK:
-        self.events.add(EventName.locationdTemporaryError)
+    #     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):
+    #   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)
+    #     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 not SIMULATION or REPLAY:
-      if self.sm['modelV2'].frameDropPerc > 1:
+    #   if self.sm['modelV2'].frameDropPerc > 1 and not self.wgpu_enabled:
-        self.events.add(EventName.modeldLagging)
+    #     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)
+    wgpu_running = ui_state.wgpu_enabled and ui_state.sm.alive["modelV2"] and ui_state.sm.valid["modelV2"]
-    self._egpu_icon_gray.set_visible(ui_state.usbgpu and not ui_state.usbgpu_compiled)
+    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/sunnypilot/layouts/settings/models.py

@@ -43,7 +43,7 @@ class ModelsLayout(Widget):
     self._initialize_items()
 
     self.clear_cache_item.action_item.set_value(f"{self.calculate_cache_size():.2f} MB")
-    for ctrl, key in [(self.lane_turn_value_control, "LaneTurnValue"), (self.delay_control, "LagdToggleDelay"), (self.camera_offset, "CameraOffset")]:
+    for ctrl, key in [(self.lane_turn_value_control, "LaneTurnValue"), (self.delay_control, "LagdToggleDelay")]:
       ctrl.action_item.set_value(int(float(ui_state.params.get(key, return_default=True)) * 100))
 
     self._scroller = Scroller(self.items, line_separator=True, spacing=0)
@@ -93,14 +93,9 @@ class ModelsLayout(Widget):
 
     self.lagd_toggle = toggle_item_sp(tr("Live Learning Steer Delay"), "", param="LagdToggle")
 
-    self.camera_offset = option_item_sp(tr("Adjust Camera Offset"), "CameraOffset", -35, 35,
-                                        tr("Virtually shift camera's perspective to move model's center to Left(+ values) or Right (- values)"),
-                                        1, None, True, "", style.BUTTON_ACTION_WIDTH, None, True,
-                                        lambda v: f"{v / 100:.2f} m")
-
     self.items = [self.current_model_item, self.cancel_download_item, self.supercombo_label, self.vision_label,
-                  self.policy_label, self.off_policy_label, self.on_policy_label, self.refresh_item, self.clear_cache_item,
+                  self.policy_label, self.off_policy_label, self.on_policy_label, self.refresh_item, self.clear_cache_item, self.lane_turn_desire_toggle,
-                  self.lane_turn_desire_toggle, self.lane_turn_value_control, self.lagd_toggle, self.delay_control, self.camera_offset]
+                  self.lane_turn_value_control, self.lagd_toggle, self.delay_control]
 
   def _update_lagd_description(self, lagd_toggle: bool):
     desc = tr("Enable this for the car to learn and adapt its steering response time. Disable to use a fixed steering response time. " +
@@ -237,7 +232,6 @@ class ModelsLayout(Widget):
     advanced_controls: bool = ui_state.params.get_bool("ShowAdvancedControls")
     turn_desire: bool = ui_state.params.get_bool("LaneTurnDesire")
     live_delay: bool = ui_state.params.get_bool("LagdToggle")
-    camera_offset: bool = ui_state.params.get("ModelManager_ActiveBundle") is not None
 
     self.lane_turn_desire_toggle.action_item.set_state(turn_desire)
     self.lane_turn_value_control.set_visible(turn_desire and advanced_controls)
@@ -246,7 +240,6 @@ class ModelsLayout(Widget):
     new_step = int(round(100 / CV.MPH_TO_KPH)) if ui_state.is_metric else 100
     if self.lane_turn_value_control.action_item is not None and self.lane_turn_value_control.action_item.value_change_step != new_step:
       self.lane_turn_value_control.action_item.value_change_step = new_step
-    self.camera_offset.set_visible(camera_offset)
 
     self._update_lagd_description(live_delay)
     self.model_manager = ui_state.sm["modelManagerSP"]

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/sunnypilot/modeld_v2/compile_modeld.py

@@ -10,291 +10,471 @@ import argparse
 import os
 import pickle
 import time
-from collections import defaultdict
 from functools import partial
+from collections import defaultdict
+
 import numpy as np
-os.environ['GMMU'] = '0'
+from tinygrad.tensor import Tensor
-
-def _patch_tinygrad_fetch_fw():
-  import hashlib
-  import pathlib
-  import zstandard
-  from tinygrad import helpers
-  _orig_fetch_fw = helpers.fetch_fw
-  def fetch_fw(path, name, sha256):
-    p = pathlib.Path(f"/lib/firmware/{path}/{name}.zst")
-    if p.is_file():
-      blob = zstandard.ZstdDecompressor().stream_reader(p.read_bytes()).read()
-      if hashlib.sha256(blob).hexdigest() == sha256:
-        return blob
-    return _orig_fetch_fw(path, name, sha256)
-  helpers.fetch_fw = fetch_fw
-_patch_tinygrad_fetch_fw()
-
-from openpilot.selfdrive.modeld.compile_modeld import NV12Frame, make_frame_prepare, sample_desire, sample_skip, shift_and_sample
-from tinygrad import dtypes
 from tinygrad.device import Device
 from tinygrad.engine.jit import TinyJit
-from tinygrad.tensor import Tensor
+
+from openpilot.selfdrive.modeld.compile_modeld import (
+  NV12Frame, make_frame_prepare,
+  shift_and_sample, sample_skip, sample_desire,
+)
 
 MODEL_TYPES = ('vision_policy', 'supercombo', 'vision_multi_policy')
 
 
-def _detect_desire_key(shapes: dict) -> str | None:
+def _detect_desire_key(policy_input_shapes):
-  return next((key for key in shapes if key.startswith('desire')), None)
+  for k in policy_input_shapes:
+    if k.startswith('desire'):
+      return k
+  return None
 
 
-def _detect_vision_keys(shapes: dict) -> tuple[str | None, str | None]:
+def _detect_vision_keys(vision_input_shapes):
-  img_keys = sorted(key for key in shapes if 'img' in key)
+  img_keys = sorted([k for k in vision_input_shapes if 'img' in k])
-  return (
+  road_key = next((k for k in img_keys if 'big' not in k), None)
-    next((key for key in img_keys if 'big' not in key), None),
+  wide_key = next((k for k in img_keys if 'big' in k), None)
-    next((key for key in img_keys if 'big' in key), None)
+  if road_key is None or wide_key is None:
-  )
+    raise ValueError(f"Cannot determine road/wide image keys from {list(vision_input_shapes.keys())}")
+  return road_key, wide_key
 
 
-def derive_frame_skip(vision_input_shapes: dict, policy_input_shapes: dict) -> int:
+def make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, device):
-  features_buffer = policy_input_shapes.get('features_buffer')
+  road_key, _ = _detect_vision_keys(vision_input_shapes)
-  return 1 if not features_buffer or features_buffer[1] >= 99 else 4
+  img = vision_input_shapes[road_key]
+  n_frames = img[1] // 6
+  img_buf_shape = (frame_skip * (n_frames - 1) + 1, 6, img[2], img[3])
 
+  fb = policy_input_shapes['features_buffer']
+  desire_key = _detect_desire_key(policy_input_shapes)
+  dp = policy_input_shapes[desire_key]
+  tc = policy_input_shapes.get('traffic_convention', (1, 2))
 
-def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+  npy = {
-  road_key, _ = _detect_vision_keys(input_shapes)
+    'desire': np.zeros(dp[2], dtype=np.float32),
-  if not road_key:
+    'traffic_convention': np.zeros(tc, dtype=np.float32),
-    raise ValueError("Vision road key missing from input shapes.")
-
-  img_shape = input_shapes[road_key]
-  n_frames = img_shape[1] // 6
-  img_buf_shape = (frame_skip * (n_frames - 1) + 1, 6, img_shape[2], img_shape[3])
-
-  desire_key = _detect_desire_key(input_shapes)
-  if not desire_key:
-    raise ValueError("Desire key missing from input shapes.")
-
-  desire_shape = input_shapes[desire_key]
-  features_buffer = input_shapes.get('features_buffer')
-
-  npy_arrays = {
-    'desire': np.zeros(desire_shape[2], dtype=np.float32),
     'tfm': np.zeros((3, 3), dtype=np.float32),
-    'big_tfm': np.zeros((3, 3), dtype=np.float32)
+    'big_tfm': np.zeros((3, 3), dtype=np.float32),
   }
 
-  for key, shape in input_shapes.items():
+  handled = {'features_buffer', desire_key, 'traffic_convention'}
-    if key not in npy_arrays and 'img' not in key and key not in ('features_buffer', desire_key):
+  for key, shape in policy_input_shapes.items():
-      npy_arrays[key] = np.zeros(shape, dtype=np.float32)
+    if key in handled:
+      continue
+    npy[key] = np.zeros(shape, dtype=np.float32)
 
-  queues = {
+  input_queues = {
     'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
     'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
-    'desire_q': Tensor(np.zeros((frame_skip * desire_shape[1], desire_shape[0], desire_shape[2]),
+    'feat_q': Tensor(np.zeros((frame_skip * (fb[1] - 1) + 1, fb[0], fb[2]), dtype=np.float32), device=device).contiguous().realize(),
-                  dtype=np.float32), device=device).contiguous().realize()
+    'desire_q': Tensor(np.zeros((frame_skip * dp[1], dp[0], dp[2]), dtype=np.float32), device=device).contiguous().realize(),
+    **{k: Tensor(v, device='NPY').realize() for k, v in npy.items()},
   }
-
+  return input_queues, npy
-  if features_buffer:
-    queues['feat_q'] = Tensor(np.zeros((frame_skip * (features_buffer[1] - 1) + 1, features_buffer[0], features_buffer[2]),
-                       dtype=np.float32), device=device).contiguous().realize()
-
-  queues.update({key: Tensor(value, device='NPY').realize() for key, value in npy_arrays.items()})
-  return queues, npy_arrays
 
 
-def make_split_input_queues(vision_input_shapes: dict, policy_input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+def make_run_split_policy(vision_runner, policy_runner, nv12: NV12Frame, model_w, model_h,
-  return generate_queues_and_npy({**vision_input_shapes, **policy_input_shapes}, frame_skip, device)
+                          vision_features_slice, frame_skip, desire_key, extra_policy_keys,
-
+                          vision_road_key, vision_wide_key, prepare_only=False):
-
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
-def make_supercombo_input_queues(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
-  return generate_queues_and_npy(input_shapes, frame_skip, device)
-
-
-def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, model_size: tuple[int, int],
-            features_slice: slice, frame_skip: int, input_shapes: dict, prepare_only: bool):
-  frame_prepare = make_frame_prepare(nv12, *model_size)
   sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
   sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
 
-  desire_key = _detect_desire_key(input_shapes)
+  def run_policy(img_q, big_img_q, feat_q, desire_q, desire, traffic_convention, tfm, big_tfm, frame, big_frame, **extra):
-  road_key, wide_key = _detect_vision_keys(input_shapes)
+    npy_tensors = [tfm.to(Device.DEFAULT), big_tfm.to(Device.DEFAULT),
+                   desire.to(Device.DEFAULT), traffic_convention.to(Device.DEFAULT)]
+    extra_device = {k: extra[k].to(Device.DEFAULT) for k in extra_policy_keys}
+    Tensor.realize(*npy_tensors, *extra_device.values())
+    tfm, big_tfm, desire, traffic_convention = npy_tensors
 
-  if not desire_key or not road_key or not wide_key:
+    img = shift_and_sample(img_q, frame_prepare(frame, tfm).unsqueeze(0), sample_skip_fn)
-    raise ValueError("Missing required vision or desire keys in input shapes.")
+    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm).unsqueeze(0), sample_skip_fn)
-
-  extra_keys = [key for key in input_shapes if key not in (desire_key, 'features_buffer', 'traffic_convention') and 'img' not in key]
-
-  def runner(img_q, big_img_q, feat_q, frame, big_frame, tfm, big_tfm, **kwargs):
-    desire_q = kwargs['desire_q']
-    desire = kwargs['desire']
-    traffic_convention = kwargs.get('traffic_convention')
-
-    npys = [tfm.to(Device.DEFAULT), big_tfm.to(Device.DEFAULT), desire.to(Device.DEFAULT)]
-    if traffic_convention is not None:
-      npys.append(traffic_convention.to(Device.DEFAULT))
-
-    extra_tensors = {key: kwargs[key].to(Device.DEFAULT) for key in extra_keys if key in kwargs}
-    Tensor.realize(*npys, *extra_tensors.values())
-
-    tfm_dev, big_tfm_dev, desire_dev = npys[:3]
-    traffic_conv_dev = npys[3] if traffic_convention is not None else None
-
-    img = shift_and_sample(img_q, frame_prepare(frame, tfm_dev).unsqueeze(0), sample_skip_fn)
-    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm_dev).unsqueeze(0), sample_skip_fn)
 
     if prepare_only:
       return img, big_img
 
-    desire_buf = shift_and_sample(desire_q, desire_dev.reshape(1, 1, -1), sample_desire_fn)
+    vision_out = next(iter(vision_runner({vision_road_key: img, vision_wide_key: big_img}).values())).cast('float32')
-    inputs = {desire_key: desire_buf, **extra_tensors}
 
-    if traffic_conv_dev is not None:
+    new_feat = vision_out[:, vision_features_slice].reshape(1, -1).unsqueeze(0)
-      inputs['traffic_convention'] = traffic_conv_dev
+    feat_buf = shift_and_sample(feat_q, new_feat, sample_skip_fn)
+    desire_buf = shift_and_sample(desire_q, desire.reshape(1, 1, -1), sample_desire_fn)
 
-    if vision_runner:
+    inputs = {'features_buffer': feat_buf, desire_key: desire_buf, 'traffic_convention': traffic_convention, **extra_device}
-      vision_out = next(iter(vision_runner({road_key: img, wide_key: big_img}).values()))
+    policy_out = next(iter(policy_runner(inputs).values())).cast('float32')
-      vision_out_cast = vision_out.cast('float32')
-      new_feat = vision_out_cast[:, features_slice].reshape(1, -1).unsqueeze(0)
-      inputs['features_buffer'] = shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
-      policy_outs = [next(iter(pol_runner(inputs).values())).cast('float32') for pol_runner in policy_runners]
-      return (vision_out_cast, *policy_outs) if len(policy_outs) > 1 else (vision_out_cast, policy_outs[0])
-    inputs.update({road_key: img, wide_key: big_img, 'features_buffer': sample_skip_fn(feat_q)})
-    policy_out = next(iter(policy_runners[0](inputs).values())).cast('float32')
-    new_feat = policy_out[:, features_slice].reshape(1, -1).unsqueeze(0)
-    shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
-    return policy_out
 
-  return runner
+    return vision_out, policy_out
+  return run_policy
 
 
-def compile_and_warmup(nv12: NV12Frame, model_size: tuple[int, int], prepare_only: bool, frame_skip: int, vision_runner, policy_runners: list, metadata: dict):
+def compile_split_policy(nv12: NV12Frame, model_w, model_h, prepare_only, frame_skip,
-  print(f"Compiling combined JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
+                         vision_runner, policy_runner, vision_metadata, policy_metadata):
+  print(f"Compiling combined policy JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
 
-  all_shapes = {key: value for meta in metadata.values() for key, value in meta['input_shapes'].items()}
+  vision_features_slice = vision_metadata['output_slices']['hidden_state']
+  vision_input_shapes = vision_metadata['input_shapes']
+  policy_input_shapes = policy_metadata['input_shapes']
+  desire_key = _detect_desire_key(policy_input_shapes)
+  extra_policy_keys = [k for k in policy_input_shapes if k not in ('features_buffer', desire_key, 'traffic_convention')]
+  vision_road_key, vision_wide_key = _detect_vision_keys(vision_input_shapes)
 
-  feat_meta = metadata.get('vision') or metadata.get('model') or metadata.get('policy')
+  _run = make_run_split_policy(vision_runner, policy_runner, nv12, model_w, model_h,
-  if not feat_meta:
+                               vision_features_slice, frame_skip, desire_key, extra_policy_keys,
-    raise ValueError("Could not find vision, model, or policy metadata.")
+                               vision_road_key, vision_wide_key, prepare_only)
+  run_policy_jit = TinyJit(_run, prune=True)
 
-  features_slice = feat_meta['output_slices']['hidden_state']
+  SEED = 42
-  WARP_DEV = 'CPU' if "USBGPU" in os.environ else Device.DEFAULT
 
-  run_func = create_jit_runner(vision_runner, policy_runners, nv12, model_size, features_slice, frame_skip, all_shapes, prepare_only)
+  def random_inputs_run_fn(fn, seed, test_val=None, test_buffers=None, expect_match=True):
-  run_jit = TinyJit(run_func, prune=True)
+    input_queues, npy = make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, Device.DEFAULT)
-  queues, npy_arrays = generate_queues_and_npy(all_shapes, frame_skip, Device.DEFAULT)
+    rng = np.random.default_rng(seed)
+    Tensor.manual_seed(seed)
 
-  for i in range(3):
+    testing = test_val is not None or test_buffers is not None
-    np.random.seed(42 + i)
+    n_runs = 1 if testing else 3
-    frame = Tensor.randint(nv12.size, low=0, high=256, dtype=dtypes.uint8, device=WARP_DEV).realize()
-    big_frame = Tensor.randint(nv12.size, low=0, high=256, dtype=dtypes.uint8, device=WARP_DEV).realize()
-    for arr in npy_arrays.values():
-      arr[:] = np.random.randn(*arr.shape).astype(arr.dtype)
 
-    Device.default.synchronize()
+    for i in range(n_runs):
-    start_time = time.perf_counter()
+      frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
-    run_jit(**queues, frame=frame, big_frame=big_frame)
+      big_frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
-    mid_time = time.perf_counter()
+      for v in npy.values():
-    Device.default.synchronize()
+        v[:] = rng.standard_normal(v.shape).astype(v.dtype)
-    print(f"  [{i + 1}/3] enqueue {(mid_time - start_time) * 1e3:6.2f} ms -- total {(time.perf_counter() - start_time) * 1e3:6.2f} ms")
+      Device.default.synchronize()
+      st = time.perf_counter()
+      outs = fn(**input_queues, frame=frame, big_frame=big_frame)
+      mt = time.perf_counter()
+      Device.default.synchronize()
+      et = time.perf_counter()
+      print(f"  [{i+1}/{n_runs}] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
 
-  return pickle.loads(pickle.dumps(run_jit)) if not prepare_only else run_jit
+      if i == 0:
+        val = [np.copy(v.numpy()) for v in outs]
+        buffers = [np.copy(v.numpy().copy()) for v in input_queues.values()]
+
+    if test_val is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(val, test_val, strict=True))
+      assert match == expect_match, f"outputs {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    if test_buffers is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(buffers, test_buffers, strict=True))
+      assert match == expect_match, f"buffers {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    return fn, val, buffers
+
+  print('capture + replay')
+  run_policy_jit, test_val, test_buffers = random_inputs_run_fn(run_policy_jit, SEED)
+
+  print('pickle round trip')
+  run_policy_jit = pickle.loads(pickle.dumps(run_policy_jit))
+  random_inputs_run_fn(run_policy_jit, SEED, test_val, test_buffers, expect_match=True)
+  random_inputs_run_fn(run_policy_jit, SEED+1, test_val, test_buffers, expect_match=False)
+  return run_policy_jit
 
 
-def _parse_size(size_str: str) -> tuple[int, int]:
+def derive_frame_skip(vision_input_shapes, policy_input_shapes):
-  width, height = size_str.lower().split('x')
+  fb = policy_input_shapes.get('features_buffer')
-  return int(width), int(height)
+  if fb is None:
+    return 1
+  fb_history = fb[1]
+  if fb_history >= 99:
+    return 1
+  return 4
 
 
-def read_file_chunked_to_shm(path):
+def make_supercombo_input_queues(input_shapes, frame_skip, device):
-  if not path:
+  img_shape = input_shapes.get('img', input_shapes.get('input_imgs'))
-    return None
+  if img_shape is None:
-  import atexit
+    raise ValueError("No img input found in model shapes")
-  from openpilot.common.file_chunker import read_file_chunked
-  from openpilot.system.hardware.hw import Paths
-  shm_path = os.path.join(Paths.shm_path(), os.path.basename(path))
-  atexit.register(lambda: os.path.exists(shm_path) and os.remove(shm_path))
-  with open(shm_path, 'wb') as f:
-    f.write(read_file_chunked(path))
-  return shm_path
 
+  n_frames = img_shape[1] // 6
+  img_buf_shape = (frame_skip * (n_frames - 1) + 1, 6, img_shape[2], img_shape[3])
 
-def _compile_for_resolutions(camera_resolutions: list, model_size: tuple[int, int], frame_skip: int,
+  numpy_keys = {}
-               vision_runner, policy_runners: list, metadata: dict) -> dict:
+  queue_keys = {}
-  from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+
-  return {
+  for key, shape in input_shapes.items():
-    (cam_w, cam_h): {
+    if 'img' in key:
-      name: compile_and_warmup(NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h)), model_size, prepare_only,
+      continue
-                   frame_skip, vision_runner, policy_runners, metadata)
+    if len(shape) == 3 and shape[1] > 1:
-      for name, prepare_only in [('warp_enqueue', True), ('run_policy', False)]
+      if key.startswith('desire'):
-    }
+        numpy_keys[key] = np.zeros(shape[2], dtype=np.float32)
-    for cam_w, cam_h in camera_resolutions
+        queue_keys[f'{key}_q'] = Tensor(
+          np.zeros((frame_skip * shape[1], shape[0], shape[2]), dtype=np.float32),
+          device=device).contiguous().realize()
+      elif key == 'features_buffer':
+        queue_keys['feat_q'] = Tensor(
+          np.zeros((frame_skip * (shape[1] - 1) + 1, shape[0], shape[2]), dtype=np.float32),
+          device=device).contiguous().realize()
+      else:
+        numpy_keys[key] = np.zeros(shape, dtype=np.float32)
+    elif len(shape) == 2:
+      numpy_keys[key] = np.zeros(shape, dtype=np.float32)
+
+  if 'traffic_convention' not in numpy_keys:
+    tc_shape = input_shapes.get('traffic_convention', (1, 2))
+    numpy_keys['traffic_convention'] = np.zeros(tc_shape, dtype=np.float32)
+
+  numpy_keys['tfm'] = np.zeros((3, 3), dtype=np.float32)
+  numpy_keys['big_tfm'] = np.zeros((3, 3), dtype=np.float32)
+
+  input_queues = {
+    'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    **queue_keys,
+    **{k: Tensor(v, device='NPY').realize() for k, v in numpy_keys.items()},
   }
+  return input_queues, numpy_keys
 
 
-def _load_policy_runners(args: argparse.Namespace) -> tuple[list, list]:
+def make_run_supercombo(model_runner, nv12: NV12Frame, model_w, model_h,
-  runners, keys = [], []
+                        features_slice, frame_skip, input_shapes, prepare_only=False):
-  for name, onnx_arg in [('policy', args.policy_onnx), ('off_policy', args.off_policy_onnx), ('on_policy', args.on_policy_onnx)]:
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
-    if onnx_arg:
+  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
-      runners.append(OnnxRunner(onnx_arg))
+  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
-      keys.append(name)
+
-  return runners, keys
+  desire_key = _detect_desire_key(input_shapes)
+  if desire_key is None:
+    raise ValueError(f"No desire* key found in input_shapes: {list(input_shapes.keys())}")
+  road_img_key, wide_img_key = _detect_vision_keys(input_shapes)
+  extra_policy_keys = [k for k in input_shapes
+                       if k not in (desire_key, 'features_buffer', 'traffic_convention')
+                       and 'img' not in k]
+
+  def run_supercombo(img_q, big_img_q, feat_q, desire_q,
+                     frame, big_frame, **kwargs):
+    desire = kwargs.get(desire_key)
+    traffic_convention = kwargs.get('traffic_convention')
+    tfm = kwargs['tfm']
+    big_tfm = kwargs['big_tfm']
+
+    tfm = tfm.to(Device.DEFAULT)
+    big_tfm = big_tfm.to(Device.DEFAULT)
+    desire = desire.to(Device.DEFAULT)
+    traffic_convention = traffic_convention.to(Device.DEFAULT)
+    Tensor.realize(tfm, big_tfm, desire, traffic_convention)
+
+    img = shift_and_sample(img_q, frame_prepare(frame, tfm).unsqueeze(0), sample_skip_fn)
+    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm).unsqueeze(0), sample_skip_fn)
+
+    if prepare_only:
+      return img, big_img
+
+    desire_buf = shift_and_sample(desire_q, desire.reshape(1, 1, -1), sample_desire_fn)
+    feat_buf = sample_skip_fn(feat_q)
+
+    inputs = {road_img_key: img, wide_img_key: big_img,
+              desire_key: desire_buf, 'features_buffer': feat_buf,
+              'traffic_convention': traffic_convention}
+    for k in extra_policy_keys:
+      if k in kwargs:
+        inputs[k] = kwargs[k].to(Device.DEFAULT)
+
+    model_out = next(iter(model_runner(inputs).values())).cast('float32')
+
+    new_feat = model_out[:, features_slice].reshape(1, -1).unsqueeze(0)
+    shift_and_sample(feat_q, new_feat, sample_skip_fn)
+
+    return model_out
+
+  return run_supercombo
+
+
+def make_run_vision_multi_policy(vision_runner, policy_runners, nv12: NV12Frame, model_w, model_h,
+                                 vision_features_slice, frame_skip, desire_key, extra_policy_keys,
+                                 vision_road_key, vision_wide_key, prepare_only=False):
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
+  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
+  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
+
+  def run_multi_policy(img_q, big_img_q, feat_q, desire_q, desire,
+                       traffic_convention, tfm, big_tfm, frame, big_frame, **extra):
+    npy_tensors = [tfm.to(Device.DEFAULT), big_tfm.to(Device.DEFAULT),
+                   desire.to(Device.DEFAULT), traffic_convention.to(Device.DEFAULT)]
+    extra_device = {k: extra[k].to(Device.DEFAULT) for k in extra_policy_keys}
+    Tensor.realize(*npy_tensors, *extra_device.values())
+    tfm, big_tfm, desire, traffic_convention = npy_tensors
+
+    img = shift_and_sample(img_q, frame_prepare(frame, tfm).unsqueeze(0), sample_skip_fn)
+    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm).unsqueeze(0), sample_skip_fn)
+
+    if prepare_only:
+      return img, big_img
+
+    vision_out = next(iter(vision_runner({vision_road_key: img, vision_wide_key: big_img}).values())).cast('float32')
+
+    new_feat = vision_out[:, vision_features_slice].reshape(1, -1).unsqueeze(0)
+    feat_buf = shift_and_sample(feat_q, new_feat, sample_skip_fn)
+    desire_buf = shift_and_sample(desire_q, desire.reshape(1, 1, -1), sample_desire_fn)
+
+    inputs = {'features_buffer': feat_buf, desire_key: desire_buf, 'traffic_convention': traffic_convention, **extra_device}
+
+    policy_outputs = []
+    for runner in policy_runners:
+      policy_out = next(iter(runner(inputs).values())).cast('float32')
+      policy_outputs.append(policy_out)
+
+    return (vision_out, *policy_outputs)
+
+  return run_multi_policy
+
+
+def _warmup_and_serialize(run_jit, input_queues, npy, nv12):
+  for i in range(3):
+    rng = np.random.default_rng(42 + i)
+    frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
+    big_frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
+    for v in npy.values():
+      v[:] = rng.standard_normal(v.shape).astype(v.dtype)
+    Device.default.synchronize()
+    st = time.perf_counter()
+    run_jit(**input_queues, frame=frame, big_frame=big_frame)
+    mt = time.perf_counter()
+    Device.default.synchronize()
+    et = time.perf_counter()
+    print(f"  [{i + 1}/3] enqueue {(mt - st) * 1e3:6.2f} ms -- total {(et - st) * 1e3:6.2f} ms")
+  return pickle.loads(pickle.dumps(run_jit))
+
+
+def compile_supercombo(nv12: NV12Frame, model_w, model_h, prepare_only, frame_skip,
+                       model_runner, metadata):
+  print(f"Compiling combined supercombo JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
+
+  features_slice = metadata['output_slices']['hidden_state']
+  input_shapes = metadata['input_shapes']
+
+  _run = make_run_supercombo(model_runner, nv12, model_w, model_h,
+                             features_slice, frame_skip, input_shapes, prepare_only)
+  run_jit = TinyJit(_run, prune=True)
+
+  input_queues, npy = make_supercombo_input_queues(input_shapes, frame_skip, Device.DEFAULT)
+
+  run_jit = _warmup_and_serialize(run_jit, input_queues, npy, nv12)
+  return run_jit
+
+
+def compile_multi_policy(nv12: NV12Frame, model_w, model_h, prepare_only, frame_skip,
+                         vision_runner, policy_runners, vision_metadata, policy_metadata):
+  print(f"Compiling combined multi-policy JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
+
+  vision_features_slice = vision_metadata['output_slices']['hidden_state']
+  vision_input_shapes = vision_metadata['input_shapes']
+  policy_input_shapes = policy_metadata['input_shapes']
+  desire_key = _detect_desire_key(policy_input_shapes)
+  extra_policy_keys = [k for k in policy_input_shapes if k not in ('features_buffer', desire_key, 'traffic_convention')]
+  vision_road_key, vision_wide_key = _detect_vision_keys(vision_input_shapes)
+
+  _run = make_run_vision_multi_policy(vision_runner, policy_runners, nv12, model_w, model_h,
+                                      vision_features_slice, frame_skip, desire_key, extra_policy_keys,
+                                      vision_road_key, vision_wide_key, prepare_only)
+  run_jit = TinyJit(_run, prune=True)
+
+  input_queues, npy = make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, Device.DEFAULT)
+
+  run_jit = _warmup_and_serialize(run_jit, input_queues, npy, nv12)
+  return run_jit
+
+
+def _parse_size(s):
+  w, h = s.lower().split('x')
+  return int(w), int(h)
 
 
 if __name__ == "__main__":
-  from openpilot.selfdrive.modeld.get_model_metadata import make_metadata_dict
   from tinygrad.nn.onnx import OnnxRunner
+  from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+  from openpilot.selfdrive.modeld.get_model_metadata import make_metadata_dict
 
-  parser = argparse.ArgumentParser(description="Compile combined JIT pkl for sunnypilot modeld_v2")
+  p = argparse.ArgumentParser(description="Compile combined JIT pkl for sunnypilot modeld_v2")
-  parser.add_argument('--model-type', choices=MODEL_TYPES, required=True)
+  p.add_argument('--model-type', choices=MODEL_TYPES, required=True)
-  parser.add_argument('--model-size', type=_parse_size, required=True, help='model input WxH')
+  p.add_argument('--model-size', type=_parse_size, required=True, help='model input WxH')
-  parser.add_argument('--camera-resolutions', type=_parse_size, nargs='+', required=True)
+  p.add_argument('--camera-resolutions', type=_parse_size, nargs='+', required=True)
-  parser.add_argument('--frame-skip', type=int, default=None, help='frame skip value (auto-derived if not provided)')
+  p.add_argument('--frame-skip', type=int, default=None, help='frame skip value (auto-derived if not provided)')
-  parser.add_argument('--output', required=True)
+  p.add_argument('--output', required=True)
 
-  parser.add_argument('--vision-onnx', help='vision ONNX (for split models)')
+  p.add_argument('--vision-onnx', help='vision ONNX (for split models)')
-  parser.add_argument('--policy-onnx', help='policy ONNX (for vision_policy)')
+  p.add_argument('--policy-onnx', help='policy ONNX (for vision_policy)')
-  parser.add_argument('--off-policy-onnx', help='off-policy ONNX (for vision_multi_policy)')
+  p.add_argument('--off-policy-onnx', help='off-policy ONNX (for vision_multi_policy)')
-  parser.add_argument('--on-policy-onnx', help='on-policy ONNX (for vision_multi_policy)')
+  p.add_argument('--on-policy-onnx', help='on-policy ONNX (for vision_multi_policy)')
-  parser.add_argument('--supercombo-onnx', help='supercombo ONNX (for supercombo)')
+  p.add_argument('--supercombo-onnx', help='supercombo ONNX (for supercombo)')
 
-  args = parser.parse_args()
+  args = p.parse_args()
-  output_data = defaultdict(dict)
+  out = defaultdict(dict)
-
-  args.vision_onnx = read_file_chunked_to_shm(args.vision_onnx)
-  args.policy_onnx = read_file_chunked_to_shm(args.policy_onnx)
-  args.off_policy_onnx = read_file_chunked_to_shm(args.off_policy_onnx)
-  args.on_policy_onnx = read_file_chunked_to_shm(args.on_policy_onnx)
-  args.supercombo_onnx = read_file_chunked_to_shm(args.supercombo_onnx)
-
-  vision_runner = OnnxRunner(args.vision_onnx) if args.vision_onnx else None
 
   if args.model_type == 'vision_policy':
-    assert vision_runner and args.policy_onnx
+    assert args.vision_onnx and args.policy_onnx
-    policy_runners = [OnnxRunner(args.policy_onnx)]
+    vision_runner = OnnxRunner(args.vision_onnx)
-    output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx), 'policy': make_metadata_dict(args.policy_onnx)}
+    policy_runner = OnnxRunner(args.policy_onnx)
+    out['metadata']['vision'] = make_metadata_dict(args.vision_onnx)
+    out['metadata']['policy'] = make_metadata_dict(args.policy_onnx)
+
+    frame_skip = args.frame_skip if args.frame_skip is not None else derive_frame_skip(out['metadata']['vision']['input_shapes'],
+                                                                                       out['metadata']['policy']['input_shapes'])
+
+    for cam_w, cam_h in args.camera_resolutions:
+      nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+      model_w, model_h = args.model_size
+      out[(cam_w, cam_h)] = {
+        name: compile_split_policy(nv12, model_w, model_h, prepare_only, frame_skip,
+                                   vision_runner, policy_runner,
+                                   out['metadata']['vision'], out['metadata']['policy'])
+        for name, prepare_only in [('warp_enqueue', True), ('run_policy', False)]
+      }
+
   elif args.model_type == 'supercombo':
     assert args.supercombo_onnx
-    policy_runners = [OnnxRunner(args.supercombo_onnx)]
+    model_runner = OnnxRunner(args.supercombo_onnx)
-    output_data['metadata'] = {'model': make_metadata_dict(args.supercombo_onnx)}
+    out['metadata']['model'] = make_metadata_dict(args.supercombo_onnx)
+
+    frame_skip = args.frame_skip if args.frame_skip is not None else derive_frame_skip({}, out['metadata']['model']['input_shapes'])
+
+    for cam_w, cam_h in args.camera_resolutions:
+      nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+      model_w, model_h = args.model_size
+      out[(cam_w, cam_h)] = {
+        name: compile_supercombo(nv12, model_w, model_h, prepare_only, frame_skip,
+                                 model_runner, out['metadata']['model'])
+        for name, prepare_only in [('warp_enqueue', True), ('run_policy', False)]
+      }
+
   elif args.model_type == 'vision_multi_policy':
-    assert vision_runner
+    assert args.vision_onnx
-    policy_runners, policy_names = _load_policy_runners(args)
+    vision_runner = OnnxRunner(args.vision_onnx)
-    output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx)}
+    out['metadata']['vision'] = make_metadata_dict(args.vision_onnx)
-    for name in policy_names:
-      runner_arg = getattr(args, f"{name}_onnx")
-      output_data['metadata'][name] = make_metadata_dict(runner_arg)
 
-  policy_keys = [key for key in output_data['metadata'].keys() if key != 'vision']
+    policy_runners = []
-  first_policy_meta = output_data['metadata'][policy_keys[0]] if policy_keys else {}
+    policy_onnxes = []
-  vision_meta = output_data['metadata'].get('vision', {})
+    if args.policy_onnx:
+      policy_onnxes.append(('policy', args.policy_onnx))
+    if args.off_policy_onnx:
+      policy_onnxes.append(('off_policy', args.off_policy_onnx))
+    if args.on_policy_onnx:
+      policy_onnxes.append(('on_policy', args.on_policy_onnx))
 
-  derived_frame_skip = args.frame_skip or derive_frame_skip(vision_meta.get('input_shapes', {}), first_policy_meta.get('input_shapes', {}))
+    for name, onnx_path in policy_onnxes:
-  output_data.update(_compile_for_resolutions(args.camera_resolutions, args.model_size, derived_frame_skip,
+      runner = OnnxRunner(onnx_path)
-                        vision_runner, policy_runners, output_data['metadata']))
+      policy_runners.append(runner)
+      out['metadata'][name] = make_metadata_dict(onnx_path)
 
-  with open(args.output, "wb") as file:
+    first_policy_key = policy_onnxes[0][0]
-    pickle.dump(output_data, file)
+    frame_skip = args.frame_skip if args.frame_skip is not None else derive_frame_skip(out['metadata']['vision']['input_shapes'],
+                                                                                       out['metadata'][first_policy_key]['input_shapes'])
 
+    for cam_w, cam_h in args.camera_resolutions:
+      nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+      model_w, model_h = args.model_size
+      out[(cam_w, cam_h)] = {
+        name: compile_multi_policy(nv12, model_w, model_h, prepare_only, frame_skip,
+                                   vision_runner, policy_runners,
+                                   out['metadata']['vision'], out['metadata'][first_policy_key])
+        for name, prepare_only in [('warp_enqueue', True), ('run_policy', False)]
+      }
+
+  with open(args.output, "wb") as f:
+    pickle.dump(out, f)
   pkl_size = os.path.getsize(args.output)
   print(f"Saved combined JIT to {args.output} ({pkl_size / 1e6:.2f} MB)")
 
   from openpilot.common.file_chunker import chunk_file, get_chunk_targets
   chunk_targets = get_chunk_targets(args.output, pkl_size)
   chunk_file(args.output, chunk_targets)
-  print(f"Chunked into {len(chunk_targets) - 1} file(s)")
+  num_chunks = len(chunk_targets) - 1
+  print(f"Chunked into {num_chunks} file(s)")

openpilot/sunnypilot/modeld_v2/modeld.py

@@ -7,7 +7,6 @@ See the LICENSE.md file in the root directory for more details.
 """
 
 import os
-os.environ['GMMU'] = '0'
 from openpilot.common.hardware import TICI
 os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
 USBGPU = "USBGPU" in os.environ
@@ -111,8 +110,6 @@ class ModelState(ModelStateBase):
     jits = pickle.load(open_file_chunked(pkl_path))
 
     self.DEV = Device.DEFAULT
-    self.WARP_DEV = 'CPU' if USBGPU else self.DEV
-    self.QUEUE_DEV = self.DEV
 
     metadata = jits['metadata']
     if 'model' in metadata:
@@ -124,7 +121,7 @@ class ModelState(ModelStateBase):
       self._vision_input_names = [k for k in model_metadata['input_shapes'] if 'img' in k]
       from openpilot.sunnypilot.modeld_v2.compile_modeld import make_supercombo_input_queues
       frame_skip = derive_frame_skip({}, model_metadata['input_shapes'])
-      self.input_queues, self.numpy_inputs = make_supercombo_input_queues(model_metadata['input_shapes'], frame_skip, device=self.QUEUE_DEV)
+      self.input_queues, self.numpy_inputs = make_supercombo_input_queues(model_metadata['input_shapes'], frame_skip, device=self.DEV)
     else:
       vision_metadata = metadata['vision']
       policy_keys = [k for k in metadata if k != 'vision']
@@ -142,11 +139,7 @@ class ModelState(ModelStateBase):
       policy_input_shapes = first_policy_metadata['input_shapes']
       self._vision_input_names = [k for k in vision_input_shapes if 'img' in k]
       frame_skip = derive_frame_skip(vision_input_shapes, policy_input_shapes)
-      self.input_queues, self.numpy_inputs = make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, device=self.QUEUE_DEV)
+      self.input_queues, self.numpy_inputs = make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, device=self.DEV)
-
-    self._desire_key = next(key for key in self.numpy_inputs if key.startswith('desire'))
-    self._road_key = next(key for key in self._vision_input_names if 'big' not in key)
-    self._wide_key = next(key for key in self._vision_input_names if 'big' in key)
 
     from openpilot.sunnypilot.modeld_v2.parse_model_outputs_split import Parser as SplitParser
     from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser as CombinedParser
@@ -168,11 +161,12 @@ class ModelState(ModelStateBase):
 
     self._run_policy = jits[(cam_w, cam_h)]['run_policy']
     self._warp_enqueue = jits[(cam_w, cam_h)]['warp_enqueue']
-    yuv_size = self.frame_buf_params[self._road_key][3]
+    road_name = next(k for k in self._vision_input_names if 'big' not in k)
+    yuv_size = self.frame_buf_params[road_name][3]
     self._warp_enqueue(
       **self.input_queues,
-      frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize(),
+      frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.DEV).contiguous().realize(),
-      big_frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize())
+      big_frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.DEV).contiguous().realize())
 
 
   @property
@@ -185,7 +179,7 @@ class ModelState(ModelStateBase):
 
   @property
   def desire_key(self) -> str:
-    return self._desire_key
+    return next(k for k in self.numpy_inputs if k.startswith('desire'))
 
   def run(self, bufs: dict[str, VisionBuf], transforms: dict[str, np.ndarray],
                 inputs: dict[str, np.ndarray], prepare_only: bool) -> dict[str, np.ndarray] | None:
@@ -196,19 +190,19 @@ class ModelState(ModelStateBase):
       yuv_size = self.frame_buf_params[key][3]
       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._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8', device=self.DEV)
       self.full_frames[key] = self._blob_cache[cache_key]
 
     desire_key = self.desire_key
     inputs[desire_key][0] = 0
     self.numpy_inputs[desire_key][:] = np.where(inputs[desire_key] - self.prev_desire > .99, inputs[desire_key], 0)
     self.prev_desire[:] = inputs[desire_key]
-    for key in ('traffic_convention', 'lateral_control_params', 'action_t'):
+    for key in ('traffic_convention', 'lateral_control_params'):
       if key in self.numpy_inputs and key in inputs:
         self.numpy_inputs[key][:] = inputs[key]
 
-    road_key = self._road_key
+    road_key = next(n for n in bufs if 'big' not in n)
-    wide_key = self._wide_key
+    wide_key = next(n for n in bufs if 'big' in n)
     self.numpy_inputs['tfm'][:, :] = transforms[road_key].reshape(3, 3)
     self.numpy_inputs['big_tfm'][:, :] = transforms[wide_key].reshape(3, 3)
 
@@ -231,12 +225,8 @@ class ModelState(ModelStateBase):
         policy_output = raw_outputs[i + 1].numpy().flatten()
         policy_sliced = {k: policy_output[np.newaxis, v] for k, v in policy_slices.items()}
         parsed = self.parser.parse_policy_outputs(policy_sliced)
-        if ('off' in self._policy_keys[i]
+        if 'off' in self._policy_keys[i] and self._has_on_policy:
-          and self._has_on_policy
-          and any('plan' in self._policy_slices_list[j] for j, k in enumerate(self._policy_keys) if 'on' in k.lower())):
-
           parsed.pop('plan', None)
-
         outputs.update(parsed)
 
       if 'planplus' in outputs and 'plan' in outputs:
@@ -251,20 +241,13 @@ class ModelState(ModelStateBase):
 
   def get_action_from_model(self, model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
                             lat_action_t: float, long_action_t: float, v_ego: float) -> log.ModelDataV2.Action:
-    if 'action' not in model_output:
+    plan = model_output['plan'][0]
-      plan = model_output['plan'][0]
+    desired_accel, should_stop = get_accel_from_plan(plan[:, Plan.VELOCITY][:, 0], plan[:, Plan.ACCELERATION][:, 0], self.constants.T_IDXS,
-      desired_accel, should_stop = get_accel_from_plan(plan[:, Plan.VELOCITY][:, 0], plan[:, Plan.ACCELERATION][:, 0], self.constants.T_IDXS,
+                                                     action_t=long_action_t)
-                                                       action_t=long_action_t)
+    desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, self.LONG_SMOOTH_SECONDS)
-      desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, self.LONG_SMOOTH_SECONDS)
-
-      curvature_plan = (plan + (self.PLANPLUS_CONTROL - 1.0) * model_output['planplus'][0]
-                        if 'planplus' in model_output and self.PLANPLUS_CONTROL != 1.0 else plan)
-      desired_curvature = get_curvature_from_output(model_output, curvature_plan, v_ego, lat_action_t, self.mlsim)
-    else:
-      desired_accel = model_output['action'][0, 1]
-      desired_curvature = model_output['action'][0, 0] / (max(1.0, v_ego))**2
-      should_stop = (v_ego < 0.3 and desired_accel < 0.1)
 
+    curvature_plan = plan + (self.PLANPLUS_CONTROL - 1.0) * model_output['planplus'][0] if 'planplus' in model_output and self.PLANPLUS_CONTROL != 1.0 else plan
+    desired_curvature = get_curvature_from_output(model_output, curvature_plan, v_ego, lat_action_t, self.mlsim)
     if self.generation is not None and self.generation >= 10: # smooth curvature for post FOF models
       if v_ego > self.MIN_LAT_CONTROL_SPEED:
         desired_curvature = smooth_value(desired_curvature, prev_action.desiredCurvature, self.LAT_SMOOTH_SECONDS)
@@ -417,12 +400,6 @@ def main(demo=False):
 
     bufs = {name: buf_extra if 'big' in name else buf_main for name in model.vision_input_names}
     transforms = {name: model_transform_extra if 'big' in name else model_transform_main for name in model.vision_input_names}
-
-    frame_delay = DT_MDL # compensate for time passed since the frame was captured: current_time - timestamp_eof is 50ms on average
-    action_delay = DT_MDL / 2 # middle of the interval between model output (current state) and next frame (expected state)
-    lat_action_t = lat_delay + frame_delay + action_delay
-    long_action_t = long_delay + frame_delay + action_delay
-
     inputs:dict[str, np.ndarray] = {
       model.desire_key: vec_desire,
       'traffic_convention': traffic_convention,
@@ -431,9 +408,6 @@ def main(demo=False):
     if 'lateral_control_params' in model.numpy_inputs:
       inputs['lateral_control_params'] = np.array([v_ego, lat_delay], dtype=np.float32)
 
-    if 'action_t' in model.numpy_inputs:
-      inputs['action_t'] = np.array([lat_action_t, long_action_t], dtype=np.float32)
-
     mt1 = time.perf_counter()
     model_output = model.run(bufs, transforms, inputs, prepare_only)
     mt2 = time.perf_counter()
@@ -445,7 +419,7 @@ def main(demo=False):
       posenet_send = messaging.new_message('cameraOdometry')
       mdv2sp_send = messaging.new_message('modelDataV2SP')
 
-      action = model.get_action_from_model(model_output, prev_action, lat_action_t, long_action_t, v_ego)
+      action = model.get_action_from_model(model_output, prev_action, lat_delay + DT_MDL, long_delay + DT_MDL, v_ego)
       prev_action = action
       fill_model_msg(drivingdata_send, modelv2_send, model_output, action,
                      publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id,

openpilot/sunnypilot/modeld_v2/parse_model_outputs_split.py

@@ -137,8 +137,6 @@ class Parser:
                     out_shape=(SplitModelConstants.NUM_ROAD_EDGES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
     if 'sim_pose' in outs:
       self.parse_mdn('sim_pose', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.POSE_WIDTH,))
-    if 'action' in outs:
-      self.parse_mdn('action', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.ACTION_WIDTH,))
 
   def parse_vision_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
     self.parse_mdn('pose', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.POSE_WIDTH,))

openpilot/sunnypilot/modeld_v2/tests/test_combined_pkl_loader.py

@@ -72,9 +72,15 @@ class TestStockEquivalence:
     stock_queues, stock_npy = make_input_queues(stock_shapes, frame_skip, device='NPY')
 
     # TODO-SP: remove action_t skip once SP adds prerequisite for deep models (action_t input queue)
-    skip_keys = {'action_t'}
+    # prev_feat is a stock QCOM corruption workaround handled inside the SP loader's JIT path
-    assert set(state.input_queues.keys()) == set(stock_queues.keys()) - skip_keys, \
+    skip_keys = {'action_t', 'prev_feat'}
-      f"Queue keys differ: v2={set(state.input_queues.keys())}, stock={set(stock_queues.keys())}"
+    # stock packs the per-key policy inputs into packed_npy_inputs; the npy views carry the individual keys
+    stock_queue_keys = set(stock_queues.keys())
+    if 'packed_npy_inputs' in stock_queue_keys:
+      stock_queue_keys.remove('packed_npy_inputs')
+      stock_queue_keys |= set(stock_npy.keys())
+    assert set(state.input_queues.keys()) == stock_queue_keys - skip_keys, \
+      f"Queue keys differ: v2={set(state.input_queues.keys())}, stock={stock_queue_keys}"
     assert set(state.numpy_inputs.keys()) == set(stock_npy.keys()) - skip_keys, \
       f"Npy keys differ: v2={set(state.numpy_inputs.keys())}, stock={set(stock_npy.keys())}"
 

openpilot/sunnypilot/modeld_v2/tests/test_warp.py

@@ -0,0 +1,103 @@
+import os
+os.environ['DEV'] = 'CPU'
+import pytest
+import numpy as np
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.sunnypilot.modeld_v2.warp import CAMERA_CONFIGS
+from openpilot.sunnypilot.modeld_v2.warp import Warp, MODEL_W, MODEL_H
+
+VISION_NAME_PAIRS = [  # needed to account for supercombos input_imgs
+  ('img', 'big_img'),
+  ('input_imgs', 'big_input_imgs'),
+]
+
+
+class MockVisionBuf:
+  def __init__(self, w, h):
+    self.width = w
+    self.height = h
+    _, _, _, yuv_size = get_nv12_info(w, h)
+    self.data = np.zeros(yuv_size, dtype=np.uint8)
+
+
+@pytest.mark.parametrize("buffer_length", [2, 5])
+def test_warp_initialization(buffer_length):
+  warp = Warp(buffer_length)
+  assert warp.buffer_length == buffer_length
+  assert warp.img_buffer_shape == (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+
+
+@pytest.mark.parametrize("buffer_length", [2, 5])
+@pytest.mark.parametrize("cam_w, cam_h", CAMERA_CONFIGS)
+@pytest.mark.parametrize("road, wide", VISION_NAME_PAIRS)
+def test_warp_process(buffer_length, cam_w, cam_h, road, wide):
+  warp = Warp(buffer_length)
+  mock_buf = MockVisionBuf(cam_w, cam_h)
+  transform = np.eye(3, dtype=np.float32).flatten()
+  bufs = {road: mock_buf, wide: mock_buf}
+  transforms = {road: transform, wide: transform}
+
+  out = warp.process(bufs, transforms)
+  assert isinstance(out, dict)
+  assert road in out and wide in out
+  assert out[road].shape == (1, 12, MODEL_H // 2, MODEL_W // 2)
+  assert out[wide].shape == (1, 12, MODEL_H // 2, MODEL_W // 2)
+
+  key = (cam_w, cam_h)
+  assert key in warp.jit_cache
+
+  out2 = warp.process(bufs, transforms)
+  assert out2[road].shape == out[road].shape
+
+
+@pytest.mark.parametrize("road, wide", VISION_NAME_PAIRS)
+def test_warp_buffer_shift(road, wide):
+  warp = Warp(2)
+  cam_w, cam_h = CAMERA_CONFIGS[1]
+  transform = np.eye(3, dtype=np.float32).flatten()
+
+  buf1 = MockVisionBuf(cam_w, cam_h)
+  buf1.data[0] = 255
+  bufs1 = {road: buf1, wide: buf1}
+  transforms = {road: transform, wide: transform}
+  out1 = warp.process(bufs1, transforms)
+  road1 = out1[road].numpy().copy()
+
+  buf2 = MockVisionBuf(cam_w, cam_h)
+  buf2.data[0] = 128
+  bufs2 = {road: buf2, wide: buf2}
+  out2 = warp.process(bufs2, transforms)
+  assert not np.array_equal(road1, out2[road].numpy())
+
+
+@pytest.mark.parametrize("buffer_length", [2, 5])
+@pytest.mark.parametrize("road, wide", VISION_NAME_PAIRS)
+def test_warp_buffer_accumulation(buffer_length, road, wide):
+  warp = Warp(buffer_length)
+  cam_w, cam_h = CAMERA_CONFIGS[0]
+  transform = np.eye(3, dtype=np.float32).flatten()
+  transforms = {road: transform, wide: transform}
+  outputs = []
+
+  for i in range(buffer_length + 1):
+    buf = MockVisionBuf(cam_w, cam_h)
+    buf.data[:] = i * 10
+    out = warp.process({road: buf, wide: buf}, transforms)
+    outputs.append(out[road].numpy().copy())
+
+  assert warp.full_buffers['img'].shape == (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+  for i in range(1, len(outputs)):
+    assert not np.array_equal(outputs[i - 1], outputs[i])
+
+
+def test_warp_different_cameras_same_instance():
+  warp = Warp(2)
+  transform = np.eye(3, dtype=np.float32).flatten()
+
+  buf1 = MockVisionBuf(*CAMERA_CONFIGS[0])
+  warp.process({'img': buf1, 'big_img': buf1}, {'img': transform, 'big_img': transform})
+  assert len(warp.jit_cache) == 1
+
+  buf2 = MockVisionBuf(*CAMERA_CONFIGS[1])
+  warp.process({'img': buf2, 'big_img': buf2}, {'img': transform, 'big_img': transform})
+  assert len(warp.jit_cache) == 2

openpilot/sunnypilot/modeld_v2/warp.py

@@ -0,0 +1,171 @@
+import pickle
+import time
+import numpy as np
+from pathlib import Path
+from tinygrad.tensor import Tensor
+from tinygrad.engine.jit import TinyJit
+from tinygrad.device import Device
+
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+from openpilot.selfdrive.modeld.compile_modeld import NV12Frame, make_frame_prepare as _make_frame_prepare
+
+CAMERA_CONFIGS = [
+  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),
+  (_os_fisheye.width, _os_fisheye.height),
+]
+
+
+def make_frame_prepare(cam_w, cam_h, model_w, model_h):
+  nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+  return _make_frame_prepare(nv12, model_w, model_h)
+
+
+def warp_pkl_path(w, h):
+  from openpilot.selfdrive.modeld.helpers import MODELS_DIR
+  return MODELS_DIR / f'warp_{w}x{h}_tinygrad.pkl'
+
+
+def make_update_img_input(frame_prepare, model_w, model_h):
+  def update_img_input_tinygrad(tensor, frame, M_inv):
+    M_inv = M_inv.to(Device.DEFAULT)
+    new_img = frame_prepare(frame, M_inv)
+    tensor.assign(tensor[6:].cat(new_img, dim=0).contiguous())
+    return Tensor.cat(tensor[:6], tensor[-6:], dim=0).contiguous().reshape(1, 12, model_h//2, model_w//2)
+  return update_img_input_tinygrad
+
+
+def make_update_both_imgs(frame_prepare, model_w, model_h):
+  update_img = make_update_img_input(frame_prepare, model_w, model_h)
+  def update_both_imgs_tinygrad(calib_img_buffer, new_img, M_inv,
+                                calib_big_img_buffer, new_big_img, M_inv_big):
+    calib_img_pair = update_img(calib_img_buffer, new_img, M_inv)
+    calib_big_img_pair = update_img(calib_big_img_buffer, new_big_img, M_inv_big)
+    return calib_img_pair, calib_big_img_pair
+  return update_both_imgs_tinygrad
+
+MODELS_DIR = Path(__file__).parent / 'models'
+MODEL_W, MODEL_H = MEDMODEL_INPUT_SIZE
+UPSTREAM_BUFFER_LENGTH = 5
+
+
+def v2_warp_pkl_path(cam_w, cam_h, buffer_length):
+  return MODELS_DIR / f'warp_{cam_w}x{cam_h}_b{buffer_length}_tinygrad.pkl'
+
+
+def compile_v2_warp(cam_w, cam_h, buffer_length):
+  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
+  img_buffer_shape = (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+
+  print(f"Compiling v2 warp for {cam_w}x{cam_h} buffer_length={buffer_length}...")
+
+  frame_prepare = make_frame_prepare(cam_w, cam_h, MODEL_W, MODEL_H)
+  update_both_imgs = make_update_both_imgs(frame_prepare, MODEL_W, MODEL_H)
+  update_img_jit = TinyJit(update_both_imgs, prune=True)
+
+  full_buffer = Tensor.zeros(img_buffer_shape, dtype='uint8').contiguous().realize()
+  big_full_buffer = Tensor.zeros(img_buffer_shape, dtype='uint8').contiguous().realize()
+  new_frame_np = np.random.default_rng(0).integers(0, 256, yuv_size, dtype=np.uint8)
+  new_big_frame_np = np.random.default_rng(1).integers(0, 256, yuv_size, dtype=np.uint8)
+  for i in range(10):
+    img_inputs = [full_buffer,
+                  Tensor.from_blob(new_frame_np.ctypes.data, (yuv_size,), dtype='uint8').realize(),
+                  Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
+    big_img_inputs = [big_full_buffer,
+                      Tensor.from_blob(new_big_frame_np.ctypes.data, (yuv_size,), dtype='uint8').realize(),
+                      Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
+    inputs = img_inputs + big_img_inputs
+    Device.default.synchronize()
+
+    st = time.perf_counter()
+    _ = update_img_jit(*inputs)
+    mt = time.perf_counter()
+    Device.default.synchronize()
+    et = time.perf_counter()
+    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
+
+  pkl_path = v2_warp_pkl_path(cam_w, cam_h, buffer_length)
+  with open(pkl_path, "wb") as f:
+    pickle.dump(update_img_jit, f)
+  print(f"  Saved to {pkl_path}")
+
+  jit = pickle.load(open(pkl_path, "rb"))
+  verify_frame = np.random.default_rng(0).integers(0, 256, yuv_size, dtype=np.uint8)
+  verify_big_frame = np.random.default_rng(1).integers(0, 256, yuv_size, dtype=np.uint8)
+  fresh_inputs = [
+    Tensor.zeros(img_buffer_shape, dtype='uint8').contiguous().realize(),
+    Tensor.from_blob(verify_frame.ctypes.data, (yuv_size,), dtype='uint8').realize(),
+    Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY'),
+    Tensor.zeros(img_buffer_shape, dtype='uint8').contiguous().realize(),
+    Tensor.from_blob(verify_big_frame.ctypes.data, (yuv_size,), dtype='uint8').realize(),
+    Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY'),
+  ]
+  jit(*fresh_inputs)
+
+
+class Warp:
+  def __init__(self, buffer_length=2):
+    self.buffer_length = buffer_length
+    self.img_buffer_shape = (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+
+    self.jit_cache = {}
+    self.full_buffers = {k: Tensor.zeros(self.img_buffer_shape, dtype='uint8').contiguous().realize() for k in ['img', 'big_img']}
+    self._blob_cache: dict[int, Tensor] = {}
+    self._nv12_cache: dict[tuple[int, int], int] = {}
+    self.transforms_np = {k: np.zeros((3, 3), dtype=np.float32) for k in ['img', 'big_img']}
+    self.transforms = {k: Tensor(v, device='NPY').realize() for k, v in self.transforms_np.items()}
+
+  def process(self, bufs, transforms):
+    if not bufs:
+      return {}
+    road = next(n for n in bufs if 'big' not in n)
+    wide = next(n for n in bufs if 'big' in n)
+    cam_w, cam_h = bufs[road].width, bufs[road].height
+    key = (cam_w, cam_h)
+
+    if key not in self.jit_cache:
+      v2_pkl = v2_warp_pkl_path(cam_w, cam_h, self.buffer_length)
+      if v2_pkl.exists():
+        with open(v2_pkl, 'rb') as f:
+          self.jit_cache[key] = pickle.load(f)
+      elif self.buffer_length == UPSTREAM_BUFFER_LENGTH:
+        upstream_pkl = warp_pkl_path(cam_w, cam_h)
+        if upstream_pkl.exists():
+          with open(upstream_pkl, 'rb') as f:
+            self.jit_cache[key] = pickle.load(f)
+      if key not in self.jit_cache:
+        frame_prepare = make_frame_prepare(cam_w, cam_h, MODEL_W, MODEL_H)
+        update_both_imgs = make_update_both_imgs(frame_prepare, MODEL_W, MODEL_H)
+        self.jit_cache[key] = TinyJit(update_both_imgs, prune=True)
+
+    if key not in self._nv12_cache:
+      self._nv12_cache[key] = get_nv12_info(cam_w, cam_h)[3]
+    yuv_size = self._nv12_cache[key]
+
+    road_ptr = bufs[road].data.ctypes.data
+    wide_ptr = bufs[wide].data.ctypes.data
+    if road_ptr not in self._blob_cache:
+      self._blob_cache[road_ptr] = Tensor.from_blob(road_ptr, (yuv_size,), dtype='uint8')
+    if wide_ptr not in self._blob_cache:
+      self._blob_cache[wide_ptr] = Tensor.from_blob(wide_ptr, (yuv_size,), dtype='uint8')
+    road_blob = self._blob_cache[road_ptr]
+    wide_blob = self._blob_cache[wide_ptr] if wide_ptr != road_ptr else Tensor.from_blob(wide_ptr, (yuv_size,), dtype='uint8')
+    np.copyto(self.transforms_np['img'], transforms[road].reshape(3, 3))
+    np.copyto(self.transforms_np['big_img'], transforms[wide].reshape(3, 3))
+
+    Device.default.synchronize()
+    res = self.jit_cache[key](
+      self.full_buffers['img'], road_blob, self.transforms['img'],
+      self.full_buffers['big_img'], wide_blob, self.transforms['big_img'],
+    )
+    out_road = res[0].realize()
+    out_wide = res[1].realize()
+
+    return {road: out_road, wide: out_wide}
+
+
+if __name__ == "__main__":
+  for cam_w, cam_h in CAMERA_CONFIGS:
+    for bl in [2, 5]:
+      compile_v2_warp(cam_w, cam_h, bl)

openpilot/sunnypilot/models/fetcher.py

@@ -6,12 +6,11 @@ See the LICENSE.md file in the root directory for more details.
 """
 
 import time
-import os
+
 import requests
 from requests.exceptions import (SSLError, RequestException, HTTPError)
 from openpilot.common.params import Params
 from openpilot.common.swaglog import cloudlog
-from openpilot.system.hardware.hw import Paths
 from openpilot.sunnypilot.models.helpers import is_bundle_version_compatible
 
 from openpilot.cereal import custom
@@ -27,35 +26,11 @@ class ModelParser:
     download_uri.sha256 = download_uri_data.get("sha256")
     return download_uri
 
-  @staticmethod
-  def _parse_chunk(chunk_data) -> custom.ModelManagerSP.Chunk:
-    chunk = custom.ModelManagerSP.Chunk()
-    chunk.fileName = chunk_data.get("file_name")
-    chunk.sha256 = chunk_data.get("sha256")
-    return chunk
-
   @staticmethod
   def _parse_artifact(artifact_data) -> custom.ModelManagerSP.Artifact:
     artifact = custom.ModelManagerSP.Artifact()
     artifact.fileName = artifact_data.get("file_name")
     artifact.downloadUri = ModelParser._parse_download_uri(artifact_data.get("download_uri", {}))
-
-    if "chunks" in artifact_data:
-      artifact.chunks = [ModelParser._parse_chunk(chunk_data) for chunk_data in artifact_data["chunks"]]
-
-      try:
-        model_dir = Paths.model_root()
-        os.makedirs(model_dir, exist_ok=True)
-        manifest_path = os.path.join(model_dir, f"{artifact.fileName}.chunkmanifest")
-        num_chunks = str(len(artifact.chunks))
-
-        if not os.path.exists(manifest_path) or open(manifest_path).read().strip() != num_chunks:
-          with open(manifest_path, "w") as f:
-            f.write(num_chunks)
-          cloudlog.info(f"Wrote chunk manifest for {artifact.fileName}: {num_chunks} chunks")
-      except Exception as e:
-        cloudlog.warning(f"Failed to write chunk manifest for {artifact.fileName}: {e}")
-
     return artifact
 
   @staticmethod
@@ -141,7 +116,7 @@ class ModelCache:
 
 class ModelFetcher:
   """Handles fetching and caching of model data from remote source"""
-  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v18.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v17.json"
 
   def __init__(self, params: Params):
     self.params = params
@@ -209,7 +184,4 @@ if __name__ == "__main__":
       # Print artifact details
       print(f"Artifact: {model.artifact.fileName}, Download URI: {model.artifact.downloadUri.uri}")
       # Print metadata details
-      if model.artifact.chunks:
+      print(f"Metadata: {model.metadata.fileName}, Download URI: {model.metadata.downloadUri.uri}")
-        print(f"Contains {len(model.artifact.chunks)} chunks.")
-      if hasattr(model, 'metadata') and model.metadata and model.metadata.fileName:
-        print(f"Metadata: {model.metadata.fileName}, Download URI: {model.metadata.downloadUri.uri}")

openpilot/sunnypilot/models/manager.py

@@ -89,16 +89,20 @@ class ModelManagerSP:
         del self._download_start_times[model.fileName]
 
   async def _download_chunked(self, base_url: str, base_path: str, artifact) -> None:
-    from openpilot.common.file_chunker import get_chunk_name, get_manifest_path
+    from openpilot.common.file_chunker import get_manifest_path, get_chunk_name
-
+    manifest_url = get_manifest_path(base_url)
-    num_chunks = len(artifact.chunks)
-    if num_chunks == 0:
-      raise ValueError("No chunks defined in artifact")
-
     manifest_path = get_manifest_path(base_path)
+
+    async with aiohttp.ClientSession() as session:
+      async with session.get(manifest_url) as resp:
+        if resp.status == 404:
+          raise FileNotFoundError
+        resp.raise_for_status()
+        num_chunks = int((await resp.read()).strip())
+
     self._download_start_times[artifact.fileName] = time.monotonic()
 
-    for i, _ in enumerate(artifact.chunks):
+    for i in range(num_chunks):
       chunk_url = get_chunk_name(base_url, i, num_chunks)
       chunk_path = get_chunk_name(base_path, i, num_chunks)
       chunk_downloaded = 0
@@ -113,7 +117,7 @@ class ModelManagerSP:
               if self.params.get("ModelManager_DownloadIndex") is None:
                 raise Exception("Download cancelled")
               intra = chunk_downloaded / max(chunk_size, 1)
-              progress = min(99.0, ((i + intra) / num_chunks) * 100)
+              progress = min(99, (i + intra) / num_chunks * 100)
               artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloading
               artifact.downloadProgress.progress = progress
               artifact.downloadProgress.eta = self._calculate_eta(artifact.fileName, progress)
@@ -144,9 +148,9 @@ class ModelManagerSP:
         self._report_status()
         return
 
-      if len(artifact.chunks) > 0:
+      try:
         await self._download_chunked(url, full_path, artifact)
-      else:
+      except (FileNotFoundError, aiohttp.ClientResponseError):
         await self._download_file(url, full_path, artifact)
 
       if not await verify_file(full_path, expected_hash):
@@ -166,15 +170,18 @@ class ModelManagerSP:
       artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
       artifact.downloadProgress.eta = 0
       self._sync_artifact_progress(artifact)
-      if self.selected_bundle:
+      self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.failed
-        self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.failed
       self._report_status()
       self._download_start_times.pop(artifact.fileName, None)
       raise
 
   async def _process_model(self, model, destination_path: str) -> None:
     """Processes a single model download including verification"""
-    await self._process_artifact(model.artifact, destination_path)
+    model_artifact = model.artifact
+    metadata_artifact = model.metadata
+
+    await self._process_artifact(metadata_artifact, destination_path)
+    await self._process_artifact(model_artifact, destination_path)
 
   def _report_status(self) -> None:
     """Reports current status through messaging system"""

openpilot/sunnypilot/models/runners/helpers.py

@@ -0,0 +1,28 @@
+from openpilot.sunnypilot.models.helpers import get_active_bundle
+from openpilot.sunnypilot.models.runners.model_runner import ModelRunner
+from openpilot.sunnypilot.models.runners.tinygrad.tinygrad_runner import TinygradRunner, TinygradSplitRunner
+from openpilot.sunnypilot.models.runners.constants import ModelType
+
+
+def get_model_runner() -> ModelRunner:
+  """
+  Factory function to create and return the appropriate ModelRunner instance.
+
+  Selects TinygradRunner, choosing TinygradSplitRunner if separate vision/policy
+  models are detected in the active bundle.
+
+  :return: An instance of a ModelRunner subclass (ONNXRunner, TinygradRunner, or TinygradSplitRunner).
+  """
+  bundle = get_active_bundle()
+  if bundle and bundle.models:
+    model_types = {m.type.raw for m in bundle.models}
+    # Check if the bundle uses separate vision and policy models (legacy or new split format)
+    split_types = {ModelType.vision, ModelType.policy, ModelType.offPolicy, ModelType.onPolicy}
+    if model_types & split_types:
+      return TinygradSplitRunner()
+    # Otherwise, assume a single model (likely supercombo)
+    if bundle.models:
+      return TinygradRunner(bundle.models[0].type.raw)
+
+  # Default fallback to TinygradRunner with the supercombo type if bundle info is missing/incomplete
+  return TinygradRunner(ModelType.supercombo)

openpilot/sunnypilot/models/runners/model_runner.py

@@ -0,0 +1,174 @@
+from abc import abstractmethod, ABC
+
+import numpy as np
+from openpilot.sunnypilot.models.helpers import get_active_bundle
+from openpilot.sunnypilot.models.runners.constants import NumpyDict, ShapeDict, Model, SliceDict, SEND_RAW_PRED
+from openpilot.common.hardware.hw import Paths
+import pickle
+
+CUSTOM_MODEL_PATH = Paths.model_root()
+
+
+class ModelData:
+  """
+  Stores metadata and configuration for a specific machine learning model.
+
+  This class loads model metadata (like input shapes and output slices)
+  from a pickle file associated with a model instance.
+
+  :param model: The machine learning model object containing metadata.
+  """
+  def __init__(self, model: Model):
+    self.model = model
+    self.metadata = model.metadata
+    self.input_shapes: ShapeDict = {}
+    self.output_slices: SliceDict = {}
+    if self.metadata:
+      self._load_metadata()
+
+  def _load_metadata(self) -> None:
+    """Loads input shapes and output slices from the model's metadata pickle file."""
+    metadata_path = f"{CUSTOM_MODEL_PATH}/{self.metadata.fileName}"
+    with open(metadata_path, 'rb') as f:
+      model_metadata = pickle.load(f)
+    self.input_shapes = model_metadata.get('input_shapes', {})
+    self.output_slices = model_metadata.get('output_slices', {})
+
+
+class ModularRunner(ABC):
+  """
+  Represents a modular runner for handling and slicing model outputs.
+
+  This abstract base class is designed to provide an interface for modular
+  parsing and processing of model outputs. Classes inheriting from it must
+  implement the specified abstract methods, defining how model outputs
+  should be handled and stored. The primary goal is to enable structured
+  parsing of outputs through a dictionary-based method mapping.
+
+  :ivar parser_method_dict: Mapping dictionary containing parser methods
+      for handling specific types of outputs.
+  :type parser_method_dict: dict
+  """
+
+  @property
+  @abstractmethod
+  def parser_method_dict(self) -> dict:
+    pass
+
+  @parser_method_dict.setter
+  @abstractmethod
+  def parser_method_dict(self, value: dict) -> None:
+    pass
+
+  @abstractmethod
+  def _slice_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    pass
+
+
+class ModelRunner(ModularRunner):
+  """
+  Abstract base class for managing and executing machine learning models.
+
+  Provides a common interface for loading models, preparing inputs, running
+  inference, and slicing/parsing outputs based on model metadata. Derived
+  classes implement the specifics of input preparation and model execution
+  for different frameworks (e.g., Tinygrad, ONNX).
+  """
+
+  def __init__(self):
+    """Initializes the model runner, loading the active model bundle."""
+    self.is_20hz: bool | None = None
+    self.is_20hz_3d: bool | None = None
+    self.models: dict[int, ModelData] = {}
+    self._model_data: ModelData | None = None  # Active model data for current operation
+    self._parser_method_dict: dict = {}
+    self.inputs: dict = {}
+    self._parser = None
+    self._load_models()
+    self._constants = None
+
+  @property
+  def constants(self):
+    return self._constants
+
+  @property
+  def parser_method_dict(self) -> dict:
+    """Returns the dictionary mapping model types to their respective parsing methods."""
+    return self._parser_method_dict
+
+  @parser_method_dict.setter
+  def parser_method_dict(self, value: dict) -> None:
+    """Sets the dictionary mapping model types to their respective parsing methods."""
+    self._parser_method_dict = value
+
+  def _load_models(self) -> None:
+    """Loads the active model bundle configuration and sets up ModelData."""
+    bundle = get_active_bundle()
+    if not bundle:
+      raise ValueError("No active model bundle found, why are we being executed?")
+
+    self.models = {model.type.raw: ModelData(model) for model in bundle.models}
+    self.is_20hz = bundle.is20hz
+    self.is_20hz_3d = False
+
+  @property
+  def input_shapes(self) -> ShapeDict:
+    """Returns the input shapes for the currently active model."""
+    if self._model_data:
+      return self._model_data.input_shapes
+    raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
+
+  @property
+  def output_slices(self) -> SliceDict:
+    """Returns the output slices for the currently active model."""
+    if self._model_data:
+      return self._model_data.output_slices
+    raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
+
+  @property
+  def vision_input_names(self) -> list[str]:
+    """Returns the list of vision input names from the input shapes."""
+    if self._model_data:
+      return list(self._model_data.input_shapes.keys())
+    raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
+
+  @abstractmethod
+  def prepare_inputs(self, numpy_inputs: NumpyDict) -> dict:
+    """
+    Abstract method to prepare inputs for model inference.
+
+    :param numpy_inputs: Dictionary of numpy arrays for non-image inputs.
+    :return: Dictionary of prepared inputs ready for the model.
+    """
+    raise NotImplementedError
+
+  @abstractmethod
+  def _run_model(self) -> NumpyDict:
+    """
+    Abstract method to execute model inference with prepared inputs.
+
+    :return: Dictionary containing the model's raw output arrays.
+    """
+    raise NotImplementedError
+
+  def _slice_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """
+    Slices the raw model output array based on the output_slices metadata.
+
+    :param model_outputs: The raw numpy array output from the model.
+    :return: A dictionary where keys are output names and values are sliced numpy arrays.
+    """
+    if not self._model_data:
+      raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
+    sliced_outputs = {k: model_outputs[np.newaxis, v] for k, v in self._model_data.output_slices.items()}
+    if SEND_RAW_PRED:
+      sliced_outputs['raw_pred'] = model_outputs.copy()  # Optionally include the full raw output
+    return sliced_outputs
+
+  def run_model(self) -> NumpyDict:
+    """
+    Executes the model inference pipeline: runs the model and parses outputs.
+
+    :return: Dictionary containing the final parsed model outputs.
+    """
+    return self._run_model()  # Parsing is handled within specific runner implementations

openpilot/sunnypilot/models/runners/tinygrad/model_types.py

@@ -0,0 +1,91 @@
+import os
+from abc import ABC
+
+import numpy as np
+from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser as CombinedParser
+from openpilot.sunnypilot.modeld_v2.parse_model_outputs_split import Parser as SplitParser
+from openpilot.sunnypilot.models.runners.constants import ModelType, NumpyDict
+from openpilot.sunnypilot.models.runners.model_runner import ModularRunner
+from openpilot.common.hardware.hw import Paths
+
+
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+CUSTOM_MODEL_PATH = Paths.model_root()
+
+
+class OffPolicyTinygrad(ModularRunner, ABC):
+  """
+  A TinygradRunner specialized for off-policy models.
+
+  Uses a SplitParser to handle outputs specific to the off-policy part of a split model setup.
+  """
+  def __init__(self):
+    self._off_policy_parser = SplitParser()
+    self.parser_method_dict[ModelType.offPolicy] = self._parse_off_policy_outputs
+
+  def _parse_off_policy_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses off-policy model outputs using SplitParser."""
+    result: NumpyDict = self._off_policy_parser.parse_policy_outputs(self._slice_outputs(model_outputs))
+    return result
+
+
+class OnPolicyTinygrad(ModularRunner, ABC):
+  """
+  A TinygradRunner specialized for on-policy models.
+
+  Uses a SplitParser to handle outputs specific to the on-policy part of a split model setup.
+  """
+  def __init__(self):
+    self._on_policy_parser = SplitParser()
+    self.parser_method_dict[ModelType.onPolicy] = self._parse_on_policy_outputs
+
+  def _parse_on_policy_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses on-policy model outputs using SplitParser."""
+    result: NumpyDict = self._on_policy_parser.parse_policy_outputs(self._slice_outputs(model_outputs))
+    return result
+
+
+class PolicyTinygrad(ModularRunner, ABC):
+  """
+  A TinygradRunner specialized for policy-only models.
+
+  Uses a SplitParser to handle outputs specific to the policy part of a split model setup.
+  """
+  def __init__(self):
+    self._policy_parser = SplitParser()
+    self.parser_method_dict[ModelType.policy] = self._parse_policy_outputs
+
+  def _parse_policy_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses policy model outputs using SplitParser."""
+    result: NumpyDict = self._policy_parser.parse_policy_outputs(self._slice_outputs(model_outputs))
+    return result
+
+class VisionTinygrad(ModularRunner, ABC):
+  """
+  A TinygradRunner specialized for vision-only models.
+
+  Uses a SplitParser to handle outputs specific to the vision part of a split model setup.
+  """
+  def __init__(self):
+    self._vision_parser = SplitParser()
+    self.parser_method_dict[ModelType.vision] = self._parse_vision_outputs
+
+  def _parse_vision_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses vision model outputs using SplitParser."""
+    result: NumpyDict = self._vision_parser.parse_vision_outputs(self._slice_outputs(model_outputs))
+    return result
+
+class SupercomboTinygrad(ModularRunner, ABC):
+  """
+  A TinygradRunner specialized for vision-only models.
+
+  Uses a SplitParser to handle outputs specific to the vision part of a split model setup.
+  """
+  def __init__(self):
+    self._supercombo_parser = CombinedParser()
+    self.parser_method_dict[ModelType.supercombo] = self._parse_supercombo_outputs
+
+  def _parse_supercombo_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses vision model outputs using SplitParser."""
+    result: NumpyDict = self._supercombo_parser.parse_outputs(self._slice_outputs(model_outputs))
+    return result

openpilot/sunnypilot/models/runners/tinygrad/tinygrad_runner.py

@@ -0,0 +1,179 @@
+import pickle
+
+import numpy as np
+from openpilot.sunnypilot.models.runners.constants import NumpyDict, ModelType, ShapeDict, CUSTOM_MODEL_PATH, SliceDict
+from openpilot.sunnypilot.models.runners.model_runner import ModelRunner
+from openpilot.sunnypilot.models.runners.tinygrad.model_types import PolicyTinygrad, VisionTinygrad, SupercomboTinygrad, OffPolicyTinygrad, OnPolicyTinygrad
+from openpilot.sunnypilot.models.split_model_constants import SplitModelConstants
+from openpilot.sunnypilot.modeld_v2.constants import ModelConstants
+
+from tinygrad.tensor import Tensor
+
+
+class TinygradRunner(ModelRunner, SupercomboTinygrad, PolicyTinygrad, VisionTinygrad, OffPolicyTinygrad, OnPolicyTinygrad):
+  """
+  A ModelRunner implementation for executing Tinygrad models.
+
+  Handles loading Tinygrad model artifacts (.pkl), preparing inputs as Tinygrad
+  Tensors (potentially using QCOM extensions on TICI), running inference,
+  and parsing the outputs.
+
+  :param model_type: The type of model (e.g., supercombo) to load and run.
+  """
+  def __init__(self, model_type: int = ModelType.supercombo):
+    ModelRunner.__init__(self)
+    SupercomboTinygrad.__init__(self)
+    PolicyTinygrad.__init__(self)
+    VisionTinygrad.__init__(self)
+    OffPolicyTinygrad.__init__(self)
+    OnPolicyTinygrad.__init__(self)
+    self._constants = ModelConstants
+    self._model_data = self.models.get(model_type)
+    if not self._model_data or not self._model_data.model:
+      raise ValueError(f"Model data for type {model_type} not available.")
+
+    artifact_filename = self._model_data.model.artifact.fileName
+    assert artifact_filename.endswith('_tinygrad.pkl'), \
+      f"Invalid model file {artifact_filename} for TinygradRunner"
+
+    model_pkl_path = f"{CUSTOM_MODEL_PATH}/{artifact_filename}"
+    with open(model_pkl_path, "rb") as f:
+      try:
+        # Load the compiled Tinygrad model runner function
+        self.model_run = pickle.load(f)
+      except FileNotFoundError as e:
+        # Provide a helpful error message if the model was built for a different platform
+        assert "/dev/kgsl-3d0" not in str(e), "Model was built on C3 or C3X, but is being loaded on PC"
+        raise
+
+    # Map input names to their required dtype and device from the loaded model
+    self.input_to_dtype = {}
+    self.input_to_device = {}
+    for idx, name in enumerate(self.model_run.captured.expected_names):
+      info = self.model_run.captured.expected_input_info[idx]
+      self.input_to_dtype[name] = info[2]  # dtype
+      self.input_to_device[name] = info[3]  # device
+    self._policy_cached = False
+
+  @property
+  def vision_input_names(self) -> list[str]:
+    """Returns the list of vision input names from the input shapes."""
+    return [name for name in self.input_shapes.keys() if 'img' in name]
+
+
+  def prepare_policy_inputs(self, numpy_inputs: NumpyDict):
+    if not self._policy_cached:
+      for key, value in numpy_inputs.items():
+        self.inputs[key] = Tensor(value, device='NPY').realize()
+      self._policy_cached = True
+
+  def prepare_inputs(self, numpy_inputs: NumpyDict) -> dict:
+    """Prepares all vision and policy inputs for the model."""
+    self.prepare_policy_inputs(numpy_inputs)
+    for key in self.vision_input_names:
+      if key in self.inputs:
+        self.inputs[key] = self.inputs[key].cast(self.input_to_dtype[key])
+    return self.inputs
+
+  def _run_model(self) -> NumpyDict:
+    """Runs the Tinygrad model inference and parses the outputs."""
+    outputs = self.model_run(**self.inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    return self._parse_outputs(outputs)
+
+  def _parse_outputs(self, model_outputs: np.ndarray) -> NumpyDict:
+    """Parses the raw model outputs using the standard Parser."""
+    if self._model_data is None:
+      raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
+
+    result: NumpyDict = self.parser_method_dict[self._model_data.model.type.raw](model_outputs)
+    return result
+
+
+class TinygradSplitRunner(ModelRunner):
+  """
+  A ModelRunner that coordinates separate TinygradVisionRunner and TinygradPolicyRunner instances.
+
+  Manages the execution of split vision and policy models, combining their inputs and outputs.
+  """
+  def __init__(self):
+    super().__init__()
+    self.is_20hz_3d = True
+    self.vision_runner = TinygradRunner(ModelType.vision)
+    self.policy_runner = TinygradRunner(ModelType.policy) if self.models.get(ModelType.policy) else None
+    self.off_policy_runner = TinygradRunner(ModelType.offPolicy) if self.models.get(ModelType.offPolicy) else None
+    self.on_policy_runner = TinygradRunner(ModelType.onPolicy) if self.models.get(ModelType.onPolicy) else None
+    self._constants = SplitModelConstants
+
+  def _run_model(self) -> NumpyDict:
+    """Runs both vision and policy models and merges their parsed outputs."""
+    vision_output = self.vision_runner.run_model()
+    outputs = {**vision_output}
+
+    if self.policy_runner:
+      policy_output = self.policy_runner.run_model()
+      outputs.update(policy_output)
+
+    if self.off_policy_runner:
+      off_policy_output = self.off_policy_runner.run_model()
+      if self.on_policy_runner:
+        off_policy_output.pop('plan', None)
+      outputs.update(off_policy_output)
+
+    if self.on_policy_runner:
+      on_policy_output = self.on_policy_runner.run_model()
+      outputs.update(on_policy_output)
+
+    if 'planplus' in outputs and 'plan' in outputs:
+      outputs['plan'] = outputs['plan'] + outputs['planplus']
+
+    return outputs
+
+  @property
+  def vision_input_names(self) -> list[str]:
+    """Returns the list of vision input names from the vision runner."""
+    return list(self.vision_runner.vision_input_names)
+
+  @property
+  def input_shapes(self) -> ShapeDict:
+    """Returns the combined input shapes from both vision and policy models."""
+    shapes = {**self.vision_runner.input_shapes}
+    if self.policy_runner:
+      shapes.update(self.policy_runner.input_shapes)
+    if self.off_policy_runner:
+      shapes.update(self.off_policy_runner.input_shapes)
+    if self.on_policy_runner:
+      shapes.update(self.on_policy_runner.input_shapes)
+    return shapes
+
+  @property
+  def output_slices(self) -> SliceDict:
+    """Returns the combined output slices from both vision and policy models."""
+    slices = {**self.vision_runner.output_slices}
+    if self.policy_runner:
+      slices.update(self.policy_runner.output_slices)
+    if self.off_policy_runner:
+      slices.update(self.off_policy_runner.output_slices)
+    if self.on_policy_runner:
+      slices.update(self.on_policy_runner.output_slices)
+    return slices
+
+  def prepare_inputs(self, numpy_inputs: NumpyDict) -> dict:
+    """Prepares inputs for both vision and policy models."""
+    if self.policy_runner:
+      self.policy_runner.prepare_policy_inputs(numpy_inputs)
+
+    for key in self.vision_input_names:
+      if key in self.inputs:
+        self.vision_runner.inputs[key] = self.inputs[key].cast(self.vision_runner.input_to_dtype[key])
+
+    inputs = {**self.vision_runner.inputs}
+    if self.policy_runner:
+      inputs.update(self.policy_runner.inputs)
+
+    if self.off_policy_runner:
+      self.off_policy_runner.prepare_policy_inputs(numpy_inputs)
+      inputs.update(self.off_policy_runner.inputs)
+    if self.on_policy_runner:
+      self.on_policy_runner.prepare_policy_inputs(numpy_inputs)
+      inputs.update(self.on_policy_runner.inputs)
+    return inputs

openpilot/sunnypilot/models/split_model_constants.py

@@ -43,7 +43,6 @@ class SplitModelConstants:
   LANE_LINES_WIDTH = 2
   ROAD_EDGES_WIDTH = 2
   PLAN_WIDTH = 15
-  ACTION_WIDTH = 2
   DESIRE_PRED_WIDTH = 8
   LAT_PLANNER_SOLUTION_WIDTH = 4
   DESIRED_CURV_WIDTH = 1

openpilot/sunnypilot/selfdrive/controls/lib/smart_cruise_control/map_controller.py

@@ -151,8 +151,8 @@ class SmartCruiseControlMap:
         a = 0.5 * TARGET_JERK
         b = self.a_ego
         c = self.v_ego - tv
-        t_a = -1 * ((b**2 - 4 * a * c) ** 0.5 + b) / (2 * a)
+        t_a = -1 * ((b**2 - 4 * a * c) ** 0.5 + b) / 2 * a
-        t_b = ((b**2 - 4 * a * c) ** 0.5 - b) / (2 * a)
+        t_b = ((b**2 - 4 * a * c) ** 0.5 - b) / 2 * a
         if not isinstance(t_a, complex) and t_a > 0:
           t = t_a
         else:

openpilot/sunnypilot/selfdrive/controls/lib/smart_cruise_control/tests/test_map_controller.py

@@ -4,17 +4,13 @@ Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
 This file is part of sunnypilot and is licensed under the MIT License.
 See the LICENSE.md file in the root directory for more details.
 """
-import json
-import math
 import platform
 
-import pytest
-
 from openpilot.cereal import custom
 from openpilot.common.params import Params
 from openpilot.common.realtime import DT_MDL
 from openpilot.selfdrive.car.cruise import V_CRUISE_UNSET
-from openpilot.sunnypilot.selfdrive.controls.lib.smart_cruise_control.map_controller import R, SmartCruiseControlMap
+from openpilot.sunnypilot.selfdrive.controls.lib.smart_cruise_control.map_controller import SmartCruiseControlMap
 
 MapState = VisionState = custom.LongitudinalPlanSP.SmartCruiseControl.MapState
 
@@ -59,17 +55,4 @@ class TestSmartCruiseControlMap:
       self.scc_m.update(True, False, 0., 0., 0.)
     assert self.scc_m.state == VisionState.enabled
 
-  def test_moderate_curve(self):
-    # Regression: `... / 2 * a` parsed as `(.../2)*a` instead of `.../(2*a)`,
-    # making max_d ~11x too small so the moderate-curve branch never tripped.
-    # v_ego=25, a_ego=0, tv=24: fixed max_d≈45m vs buggy ≈4m at a 40m waypoint.
-    waypoint_lon_deg = (40.0 / R) * (180.0 / math.pi)
-    self.mem_params.put("LastGPSPosition", json.dumps({"latitude": 0.0, "longitude": 0.0}), block=True)
-    self.mem_params.put("MapTargetVelocities",
-                        json.dumps([{"latitude": 0.0, "longitude": waypoint_lon_deg, "velocity": 24.0}]), block=True)
-
-    self.scc_m.update(True, False, 25.0, 0.0, 30.0)
-
-    assert self.scc_m.v_target == pytest.approx(24.0)
-
   # TODO-SP: mock data from modelV2 to test other states

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"
+  sock = ZmqSubSocket(sock_name, addr)
-  messaging.reset_context()
-  sock = messaging.sub_sock(sock_name, None, addr=addr, conflate=False)
-  cnt = 0
   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)
+    msgs = sock.drain(wait_for_one=True)
-    for evt in msgs:
+    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))
+      # Preserve the device camera metadata so remote model outputs line up with
-      cnt += 1
+      # 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")
+    print("waiting for remote camera stream metadata", flush=True)
-    os.environ["ZMQ"] = "1"
+    sm = ZmqSubMaster([ENCODE_SOCKETS[s] for s in vision_streams], addr)
-    messaging.reset_context()
-    sm = messaging.SubMaster([ENCODE_SOCKETS[s] for s in vision_streams], addr=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]]
+      width, height = stream_dimensions[vst]
-      self.vipc_server.create_buffers(vst, 4, ed.width, ed.height)
+      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]]
+      width, height = stream_dimensions[vst]
-      p = multiprocessing.Process(target=decoder, args=(addr, self.vipc_server, vst, ed.width, ed.height, debug))
+      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)
+  avutil = ctypes.CDLL(_library_path("avutil", 59), mode=ctypes.RTLD_GLOBAL)
-  avcodec = ctypes.CDLL(os.path.join(ffmpeg.LIB_DIR, "libavcodec.so.61"), mode=ctypes.RTLD_GLOBAL)
+  avcodec = ctypes.CDLL(_library_path("avcodec", 61), mode=ctypes.RTLD_GLOBAL)
-  swscale = ctypes.CDLL(os.path.join(ffmpeg.LIB_DIR, "libswscale.so.8"), 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)

release/ci/model_generator.py

@@ -5,6 +5,8 @@ This file is part of sunnypilot and is licensed under the MIT License.
 See the LICENSE.md file in the root directory for more details.
 """
 
+import os
+import pickle
 import sys
 import hashlib
 import json
@@ -12,8 +14,46 @@ import re
 from pathlib import Path
 from datetime import datetime, UTC
 
+REQUIRED_OUTPUT_KEYS = frozenset({
+  "plan",
+  "lane_lines",
+  "road_edges",
+  "lead",
+  "desire_state",
+  "desire_pred",
+  "meta",
+  "lead_prob",
+  "lane_lines_prob",
+  "pose",
+  "wide_from_device_euler",
+  "road_transform",
+  "hidden_state",
+})
+OPTIONAL_OUTPUT_KEYS = frozenset({
+  "planplus",
+  "sim_pose",
+  "desired_curvature",
+})
 
-def create_short_name(full_name: str) -> str:
+
+def validate_model_outputs(metadata_paths: list[Path]) -> None:
+  combined_keys: set[str] = set()
+  for path in metadata_paths:
+    if path.stat().st_size == 0:
+      print(f"skipping empty metadata: {path}")
+      continue
+    with open(path, "rb") as f:
+      metadata = pickle.load(f)
+    combined_keys.update(metadata.get("output_slices", {}).keys())
+  missing = REQUIRED_OUTPUT_KEYS - combined_keys
+  if missing:
+    raise ValueError(f"Combined model metadata is missing required output keys: {sorted(missing)}")
+  detected_optional = sorted(OPTIONAL_OUTPUT_KEYS & combined_keys)
+  if detected_optional:
+    print(f"Optional output keys detected: {detected_optional}")
+
+
+def create_short_name(full_name):
   # Remove parentheses and extract alphanumeric words
   clean_name = re.sub(r'\([^)]*\)', '', full_name)
   words = [re.sub(r'[^a-zA-Z0-9]', '', word) for word in clean_name.split() if re.sub(r'[^a-zA-Z0-9]', '', word)]
@@ -81,41 +121,49 @@ def _rename_pkl_with_chunks(old_pkl: Path, new_pkl: Path) -> Path:
   return old_pkl.rename(new_pkl)
 
 
-def generate_chunked_model(driving_pkl: Path) -> dict:
+def generate_metadata(model_path: Path, output_dir: Path, short_name: str, driving_pkl: Path):
+  base = model_path.stem
+  metadata_file = output_dir / f"{base}_metadata.pkl"
+
+  if short_name:
+    renamed_meta = output_dir / f"{base}_{short_name.lower()}_metadata.pkl"
+    if metadata_file.exists() and not renamed_meta.exists():
+      metadata_file = metadata_file.rename(renamed_meta)
+    elif renamed_meta.exists():
+      metadata_file = renamed_meta
+
+  if not metadata_file.exists():
+    print(f"Warning: Missing metadata for {base} ({metadata_file}), skipping", file=sys.stderr)
+    return
+
   tinygrad_hash = hashlib.sha256(_read_pkl_bytes(driving_pkl)).hexdigest()
 
-  chunks_config = []
+  with open(metadata_file, 'rb') as f:
-  manifest_file = Path(f"{driving_pkl}.chunkmanifest")
+    metadata_hash = hashlib.sha256(f.read()).hexdigest()
-  if manifest_file.exists():
-    num_chunks = int(manifest_file.read_text().strip())
-    for i in range(num_chunks):
-      chunk_path = Path(f"{driving_pkl}.chunk{i + 1:02d}of{num_chunks:02d}")
-      if chunk_path.exists():
-        chunk_hash = hashlib.sha256(chunk_path.read_bytes()).hexdigest()
-        chunks_config.append({
-          "file_name": chunk_path.name,
-          "sha256": chunk_hash
-        })
 
-  artifact_data = {
+  model_type = "offPolicy" if "off_policy" in base else "onPolicy" if "on_policy" in base else base.split("_")[-1]
-    "file_name": driving_pkl.name,
+
-    "download_uri": {
+  return {
-      "url": "https://gitlab.com/sunnypilot/public/docs.sunnypilot.ai/-/raw/main/",
+    "type": model_type,
-      "sha256": tinygrad_hash
+    "artifact": {
+      "file_name": driving_pkl.name,
+      "download_uri": {
+        "url": "https://gitlab.com/sunnypilot/public/docs.sunnypilot.ai/-/raw/main/",
+        "sha256": tinygrad_hash
+      }
+    },
+    "metadata": {
+      "file_name": metadata_file.name,
+      "download_uri": {
+        "url": "https://gitlab.com/sunnypilot/public/docs.sunnypilot.ai/-/raw/main/",
+        "sha256": metadata_hash
+      }
     }
   }
 
-  if chunks_config:
-    artifact_data["chunks"] = chunks_config
 
-  return {
+def create_metadata_json(models: list, output_dir: Path, custom_name=None, short_name=None, is_20hz=False, upstream_branch="unknown"):
-    "type": "chunked",
+  metadata_json = {
-    "artifact": artifact_data,
-  }
-
-
-def create_metadata_json(models: list, output_dir: Path, custom_name=None, short_name=None, is_20hz=False, upstream_branch="unknown") -> None:
-  bundle_json = {
     "short_name": short_name,
     "display_name": custom_name or upstream_branch,
     "is_20hz": is_20hz,
@@ -131,26 +179,40 @@ def create_metadata_json(models: list, output_dir: Path, custom_name=None, short
   }
 
   # Write metadata to output_dir
-  metadata_json = {
-    "bundles": [bundle_json]
-  }
-
   with open(output_dir / "metadata.json", "w") as f:
     json.dump(metadata_json, f, indent=2)
-  print("Generated metadata.json")
+
+  print(f"Generated metadata.json with {len(models)} models.")
 
 
 if __name__ == "__main__":
   import argparse
+  import glob
 
-  parser = argparse.ArgumentParser(description="Generate metadata JSON for the compiled JIT model")
+  parser = argparse.ArgumentParser(description="Generate metadata for model files")
-  parser.add_argument("--model-dir", default="./models", help="Directory containing the model files")
+  parser.add_argument("--model-dir", default="./models", help="Directory containing ONNX model files")
   parser.add_argument("--output-dir", default="./output", help="Output directory for metadata")
   parser.add_argument("--custom-name", help="Custom display name for the model")
   parser.add_argument("--is-20hz", action="store_true", help="Whether this is a 20Hz model")
+  parser.add_argument("--validate-only", action="store_true")
   parser.add_argument("--upstream-branch", default="unknown", help="Upstream branch name")
   args = parser.parse_args()
 
+  if args.validate_only:
+    metadata_paths = glob.glob(os.path.join(args.model_dir, "*_metadata.pkl"))
+    if not metadata_paths:
+      print(f"No metadata files found in {args.model_dir}", file=sys.stderr)
+      sys.exit(1)
+    validate_model_outputs([Path(p) for p in metadata_paths])
+    print(f"Validated {len(metadata_paths)} metadata files successfully.")
+    sys.exit(0)
+
+  # Find all ONNX files in the given directory
+  model_paths = glob.glob(os.path.join(args.model_dir, "*.onnx"))
+  if not model_paths:
+    print(f"No ONNX files found in {args.model_dir}", file=sys.stderr)
+    sys.exit(1)
+
   _output_dir = Path(args.output_dir)
   _output_dir.mkdir(exist_ok=True, parents=True)
   _short_name = create_short_name(args.custom_name) if args.custom_name else None
@@ -167,5 +229,14 @@ if __name__ == "__main__":
     else:
       _driving_pkl = new_pkl
 
-  _model_metadata = generate_chunked_model(_driving_pkl)
+  _models = []
-  create_metadata_json([_model_metadata], _output_dir, args.custom_name, _short_name, args.is_20hz, args.upstream_branch)
+
+  for _model_path in model_paths:
+    _model_metadata = generate_metadata(Path(_model_path), _output_dir, _short_name, _driving_pkl)
+    if _model_metadata:
+      _models.append(_model_metadata)
+
+  if _models:
+    create_metadata_json(_models, _output_dir, args.custom_name, _short_name, args.is_20hz, args.upstream_branch)
+  else:
+    print("No models processed.", file=sys.stderr)