modeld_v2: Support eGpu

* modeld_v2: safe model validation

* fix string

* numpy

* dumb

* god use full attribute names please

* modeld_v2: refactor compile_modeld

* redundant

* CREAM AND SUGAR

* gpu stuffs

* Update fetcher.py

* Update compile_modeld.py

* Update compile_modeld.py

* Update compile_modeld.py

* Update compile_modeld.py

* summary

* Update compile_modeld.py

* Update compile_modeld.py

* Update compile_modeld.py

* i could lie say

* needed

* i could

* Update compile_modeld.py

* done done done

* bye metadata

* simplify

* deeeeep

* oopsie

* i don't know what i'm doing

* it's a supercombo

* wtf. ghostwriter

* i might be blind

* fuck. i AM blind. or dumb. or both.

* hmmmm

* read/open - what's the difference

* realize. that i don't know shit

* whatever

* fix paths

* fuckit. dynamic everything.

* simplify everything.

* fix macos cabana

* np.random.seed is deprecated, use Generator or default rng instead @nayan

* this is annoying me

* too much fluff

* no

* lint be crazy now, man i've been away a while

* back

* fix pkl loader test

* that comment was wrong, its still per model, just compiled at with the input/output shapes

* fix chunking

* dump and assign to cpu

* prev_feat in cpu to prevent corruption

* add todo-sp

* desire me?

* allow legacy

* lint

* shapey

* bye

---------

Co-authored-by: discountchubbs <alexgrant990@gmail.com>
Co-authored-by: James Vecellio-Grant <159560811+Discountchubbs@users.noreply.github.com>
Co-authored-by: Jason Wen <haibin.wen3@gmail.com>

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

@@ -34,6 +34,14 @@ on:
         required: false
         default: true
         type: boolean
+      target_hardware:
+        description: 'Hardware target to compile for'
+        required: false
+        type: choice
+        default: 'qcom'
+        options:
+          - qcom
+          - usbgpu
   workflow_dispatch:
     inputs:
       upstream_branch:
@@ -81,9 +89,17 @@ on:
         description: 'Minimum selector version'
         required: false
         type: string
+      target_hardware:
+        description: 'Hardware target to compile for'
+        required: false
+        type: choice
+        default: 'qcom'
+        options:
+          - qcom
+          - usbgpu
 env:
   RECOMPILED_DIR: recompiled${{ inputs.recompiled_dir }}
-  JSON_FILE: docs/docs/driving_models_v${{ inputs.json_version }}.json
+  JSON_FILE: docs/docs/driving_models_${{ inputs.target_hardware == 'usbgpu' && 'usbgpu_v' || 'v' }}${{ inputs.json_version }}.json
 
 jobs:
   build_model:
@@ -93,6 +109,7 @@ jobs:
       custom_name: ${{ inputs.custom_name || inputs.upstream_branch }}
       is_20hz: ${{ inputs.is_20hz }}
       artifact_suffix: ${{ inputs.artifact_suffix }}
+      target_hardware: ${{ inputs.target_hardware }}
     secrets: inherit
 
   publish_model:

.github/workflows/prebuilt.yaml

@@ -1,39 +0,0 @@
-name: prebuilt
-on:
-  schedule:
-    - cron: '0 * * * *'
-  workflow_dispatch:
-
-env:
-  DOCKER_LOGIN: docker login ghcr.io -u ${{ github.actor }} -p ${{ secrets.GITHUB_TOKEN }}
-  BUILD: release/ci/docker_build_sp.sh
-
-jobs:
-  build_prebuilt:
-    name: build prebuilt
-    runs-on: ubuntu-latest
-    if: github.repository == 'sunnypilot/sunnypilot'
-    env:
-      PUSH_IMAGE: true
-    permissions:
-      checks: read
-      contents: read
-      packages: write
-    steps:
-    - name: Wait for green check mark
-      if: ${{ github.event_name != 'workflow_dispatch' }}
-      uses: lewagon/wait-on-check-action@ccfb013c15c8afb7bf2b7c028fb74dc5a068cccc
-      with:
-        ref: master
-        wait-interval: 30
-        running-workflow-name: 'build prebuilt'
-        repo-token: ${{ secrets.GITHUB_TOKEN }}
-        check-regexp: ^((?!.*(build master-ci|create badges).*).)*$
-    - uses: actions/checkout@v6
-      with:
-        submodules: true
-    - run: git lfs pull
-    - name: Build and Push docker image
-      run: |
-        $DOCKER_LOGIN
-        eval "$BUILD"

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

@@ -191,7 +191,7 @@ jobs:
           if [ "${{ inputs.target_hardware }}" == "usbgpu" ]; then
             echo "USBGPU build"
             export USBGPU=1
-            TG_FLAGS="DEV=AMD USBGPU=1 IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+            TG_FLAGS="DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0"
             OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
           else
             echo "QCOM build"

openpilot/cereal/custom.capnp

@@ -69,6 +69,8 @@ struct LeadData {
 struct SelfdriveStateSP @0x81c2f05a394cf4af {
   mads @0 :ModularAssistiveDrivingSystem;
   intelligentCruiseButtonManagement @1 :IntelligentCruiseButtonManagement;
+  buttonsPressed @2 :UInt16;
+  buttonsReleaseToggle @3 :UInt16;
 
   enum AudibleAlert {
     none @0;

openpilot/common/params_keys.h

@@ -197,7 +197,9 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"ModelManager_DownloadIndex", {CLEAR_ON_MANAGER_START | CLEAR_ON_ONROAD_TRANSITION, INT}},
     {"ModelManager_Favs", {PERSISTENT | BACKUP, STRING}},
     {"ModelManager_LastSyncTime", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, INT, "0"}},
+    {"ModelManager_LastSyncTime_USBGPU", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, INT, "0"}},
     {"ModelManager_ModelsCache", {PERSISTENT | BACKUP, JSON}},
+    {"ModelManager_ModelsCache_USBGPU", {PERSISTENT | BACKUP, JSON}},
 
     // Neural Network Lateral Control
     {"NeuralNetworkLateralControl", {PERSISTENT | BACKUP, BOOL, "0"}},
@@ -222,6 +224,7 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"SubaruStopAndGo", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"SubaruStopAndGoManualParkingBrake", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"TeslaCoopSteering", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"TeslaMadsScreenButton", {PERSISTENT | BACKUP, INT, "0"}},
     {"ToyotaEnforceStockLongitudinal", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"ToyotaStopAndGoHack", {PERSISTENT | BACKUP, BOOL, "0"}},
 

openpilot/selfdrive/controls/lib/longitudinal_planner.py

@@ -158,7 +158,7 @@ class LongitudinalPlanner(LongitudinalPlannerSP):
   def publish(self, sm, pm):
     plan_send = messaging.new_message('longitudinalPlan')
 
-    plan_send.valid = sm.all_checks(service_list=['carState', 'controlsState', 'selfdriveState', 'radarState'])
+    plan_send.valid = sm.all_checks()
 
     longitudinalPlan = plan_send.longitudinalPlan
     longitudinalPlan.modelMonoTime = sm.logMonoTime['modelV2']

openpilot/selfdrive/controls/plannerd.py

@@ -29,19 +29,19 @@ def main():
   longitudinal_planner = LongitudinalPlanner(CP, CP_SP)
   pm = messaging.PubMaster(['longitudinalPlan', 'driverAssistance', 'longitudinalPlanSP'])
   sm = messaging.SubMaster(['carControl', 'carState', 'controlsState', 'liveParameters', 'radarState', 'modelV2', 'selfdriveState',
-                            'liveMapDataSP', 'carStateSP', gps_location_service],
+                            'liveMapDataSP', 'carStateSP', 'selfdriveStateSP', gps_location_service],
-                           poll='carState', ignore_alive=ignore_services, ignore_avg_freq=ignore_services, ignore_valid=ignore_services)
+                           poll='modelV2', ignore_alive=ignore_services, ignore_avg_freq=ignore_services, ignore_valid=ignore_services)
 
   while True:
     sm.update()
-    longitudinal_planner.sla.update_car_state(sm['carState'])
+    longitudinal_planner.sla.update_buttons(sm['selfdriveStateSP'].buttonsReleaseToggle)
     if sm.updated['modelV2']:
       longitudinal_planner.update(sm)
       longitudinal_planner.publish(sm, pm)
 
       ldw.update(sm.frame, sm['modelV2'], sm['carState'], sm['carControl'])
       msg = messaging.new_message('driverAssistance')
-      msg.valid = sm.all_checks(['carState', 'carControl', 'modelV2', 'liveParameters'])
+      msg.valid = sm.all_checks()
       msg.driverAssistance.leftLaneDeparture = ldw.left
       msg.driverAssistance.rightLaneDeparture = ldw.right
       pm.send('driverAssistance', msg)

openpilot/selfdrive/modeld/compile_modeld.py

@@ -3,7 +3,6 @@ import argparse
 import atexit
 import math
 import os
-import pickle
 import tempfile
 import time
 import shutil
@@ -13,7 +12,6 @@ from collections import namedtuple
 import numpy as np
 
 from openpilot.selfdrive.modeld.helpers import dump_oob, load_oob
-from openpilot.selfdrive.modeld.usbgpu_link import wait_usbgpu_link
 
 def _patch_tinygrad_fetch_fw():
   import hashlib
@@ -31,22 +29,6 @@ def _patch_tinygrad_fetch_fw():
   helpers.fetch_fw = fetch_fw
 _patch_tinygrad_fetch_fw()
 
-def _patch_tinygrad_buffer_reduce():
-  from tinygrad.device import Buffer
-  def __reduce_ex__(self, protocol):
-    buf = None
-    if self._base is not None:
-      return self.__class__, (self.device, self.size, self.dtype, None, None, None, 0, self.base, self.offset, self.is_allocated())
-    if self.device == "NPY":
-      return self.__class__, (self.device, self.size, self.dtype, self._buf, self.options, None, self.uop_refcount)
-    if self.is_allocated():
-      buf = bytearray(self.nbytes)
-      self.copyout(memoryview(buf))
-      if protocol >= 5:
-        buf = pickle.PickleBuffer(buf)
-    return self.__class__, (self.device, self.size, self.dtype, None, self.options, buf, self.uop_refcount)
-  Buffer.__reduce_ex__ = __reduce_ex__
-_patch_tinygrad_buffer_reduce()
 
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import Context
@@ -312,9 +294,6 @@ if __name__ == "__main__":
   p.add_argument('--frame-skip', type=int, required=True)
   args = p.parse_args()
 
-  if 'USB+AMD' in os.environ.get('DEV', ''):
-    wait_usbgpu_link()
-
   model_path = read_file_chunked_to_disk(args.onnx)
   model_w, model_h = args.model_size
 

openpilot/selfdrive/modeld/helpers.py

@@ -6,10 +6,11 @@ import struct
 import tempfile
 from pathlib import Path
 
+from openpilot.common.file_chunker import get_manifest_path
+from openpilot.common.hardware.usb import CHESTNUT_VENDOR_ID, CHESTNUT_PRODUCT_ID, USB_DEVICES_PATH
+
 MODELS_DIR = Path(__file__).resolve().parent / 'models'
 TG_INPUT_DEVICES_PATH = MODELS_DIR / 'tg_input_devices.json'
-USBGPU_VID = 0xADD1
-USBGPU_PID = 0x0001
 
 
 def get_tg_input_devices(process_name: str, usbgpu: bool):
@@ -38,22 +39,21 @@ def dump_oob(obj, f):
 def load_oob(f):
   opcodes = f.read(struct.unpack('<q', f.read(8))[0])
   def buffers():
-    prev = None
     while (h := f.read(8)):
-      if prev is not None:
+      pb = pickle.PickleBuffer(bytearray(struct.unpack('<q', h)[0]))
-        prev.release()
+      f.readinto(pb)
-      buf = bytearray(struct.unpack('<q', h)[0])
+      yield pb
-      f.readinto(buf)
-      prev = pickle.PickleBuffer(buf)
-      yield prev
   return pickle.load(io.BytesIO(opcodes), buffers=buffers())
 
 def usbgpu_present() -> bool:
-  for d in Path("/sys/bus/usb/devices").glob("*"):
+  for d in USB_DEVICES_PATH.glob("*"):
     try:
-      if int((d / "idVendor").read_text(), 16) == USBGPU_VID and \
+      usb_id = (int((d / "idVendor").read_text(), 16), int((d / "idProduct").read_text(), 16))
-          int((d / "idProduct").read_text(), 16) == USBGPU_PID:
+      if usb_id == (CHESTNUT_VENDOR_ID, CHESTNUT_PRODUCT_ID):
         return True
     except Exception:
       pass
   return False
+
+def usbgpu_compiled() -> bool:
+  return Path(get_manifest_path(modeld_pkl_path(usbgpu=True))).is_file()

openpilot/selfdrive/modeld/modeld.py

@@ -2,6 +2,7 @@
 import os
 os.environ['GMMU'] = '0' # for usbgpu fast loading, noop for qcom
 from tinygrad.tensor import Tensor
+import threading
 import time
 import numpy as np
 import openpilot.cereal.messaging as messaging
@@ -18,17 +19,13 @@ from openpilot.common.transformations.camera import DEVICE_CAMERAS
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 from openpilot.common.transformations.model import get_warp_matrix
 from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
-from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value, get_curvature_from_plan
+from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, should_stop, smooth_value, get_curvature_from_plan
 from openpilot.selfdrive.modeld.parse_model_outputs import Parser
 from openpilot.selfdrive.modeld.compile_modeld import make_input_queues, WARP_INPUTS, POLICY_INPUTS
 from openpilot.selfdrive.modeld.fill_model_msg import fill_model_msg, fill_driving_model_data, fill_pose_msg, PublishState
-from openpilot.common.file_chunker import open_file_chunked, get_manifest_path
+from openpilot.common.file_chunker import open_file_chunked
 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.helpers import usbgpu_present, usbgpu_compiled, modeld_pkl_path, get_tg_input_devices, load_oob
-from openpilot.selfdrive.modeld.usbgpu_link import wait_usbgpu_link
-
-from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
-from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
 
 PROCESS_NAME = "openpilot.selfdrive.modeld.modeld"
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
@@ -36,16 +33,17 @@ SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
 LAT_SMOOTH_SECONDS = 0.0
 LONG_SMOOTH_SECONDS = 0.3
 MIN_LAT_CONTROL_SPEED = 0.3
+BIG_MODEL_TIMEOUT = 60
 
 
 def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
                           lat_action_t: float, long_action_t: float, v_ego: float) -> log.ModelDataV2.Action:
   if 'action' not in model_output:
     plan = model_output['plan'][0]
-    desired_accel, should_stop = get_accel_from_plan(plan[:,Plan.VELOCITY][:,0],
+    desired_accel = get_accel_from_plan(plan[:,Plan.VELOCITY][:,0],
-                                                     plan[:,Plan.ACCELERATION][:,0],
+                                        plan[:,Plan.ACCELERATION][:,0],
-                                                     ModelConstants.T_IDXS,
+                                        ModelConstants.T_IDXS,
-                                                     action_t=long_action_t)
+                                        action_t=long_action_t)
     desired_curvature = get_curvature_from_plan(plan[:,Plan.T_FROM_CURRENT_EULER][:,2],
                                                 plan[:,Plan.ORIENTATION_RATE][:,2],
                                                 ModelConstants.T_IDXS,
@@ -54,7 +52,7 @@ def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.
   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)
+  stop = should_stop(v_ego, desired_accel)
   desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, LONG_SMOOTH_SECONDS)
   if v_ego > MIN_LAT_CONTROL_SPEED:
     desired_curvature = smooth_value(desired_curvature, prev_action.desiredCurvature, LAT_SMOOTH_SECONDS)
@@ -63,7 +61,7 @@ def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.
 
   return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature),
                                 desiredAcceleration=float(desired_accel),
-                                shouldStop=bool(should_stop))
+                                shouldStop=bool(stop))
 
 
 class FrameMeta:
@@ -76,12 +74,10 @@ class FrameMeta:
       self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
 
 
-class ModelState(ModelStateBase):
+class ModelState:
   prev_desire: np.ndarray  # for tracking the rising edge of the pulse
 
   def __init__(self, cam_w: int, cam_h: int, usbgpu: bool):
-    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)))
@@ -138,16 +134,24 @@ class ModelState(ModelStateBase):
       outputs_dict['raw_pred'] = model_output.copy()
     return outputs_dict
 
+  def warmup(self) -> None:
+    dummy_frames = {k: np.zeros(self.frame_buf_params[k][3], dtype=np.uint8) for k in self.vision_input_names}
+    eye = np.eye(3, dtype=np.float32)
+    dims = {'desire_pulse': ModelConstants.DESIRE_LEN, 'traffic_convention': 2, 'action_t': 2}
+    self.run(dummy_frames, dict.fromkeys(self.vision_input_names, eye), {k: np.zeros(v, dtype=np.float32) for k, v in dims.items()})
+    self.input_queues, self.npy = make_input_queues(self.input_shapes, self.frame_skip, device=self.QUEUE_DEV)
+    self.prev_desire[:] = 0
+    self.full_frames.clear()
+    self._blob_cache.clear()
+
 
 def main(demo=False):
   cloudlog.warning("modeld init")
 
-  _present = usbgpu_present()
+  USBGPU = usbgpu_present() and usbgpu_compiled()
-  _compiled = os.path.isfile(get_manifest_path(modeld_pkl_path(usbgpu=True)))
-  USBGPU = _present and _compiled
   params = Params()
-  params.put_bool("UsbGpuPresent", _present)
+  params.put_bool("UsbGpuLoading", USBGPU)
-  params.put_bool("UsbGpuCompiled", _compiled)
+  params.remove("UsbGpuActive")
 
   config_realtime_process(7, 54)
 
@@ -174,15 +178,33 @@ def main(demo=False):
   if use_extra_client:
     cloudlog.warning(f"connected extra cam with buffer size: {vipc_client_extra.buffer_len} ({vipc_client_extra.width} x {vipc_client_extra.height})")
 
-  if USBGPU:
-    wait_usbgpu_link()
   st = time.monotonic()
   cloudlog.warning("loading model")
-  model = ModelState(vipc_client_main.width, vipc_client_main.height, USBGPU)
+  model = None
+  if USBGPU:
+    big_model = None
+    def load_big():
+      nonlocal big_model
+      try:
+        m = ModelState(vipc_client_main.width, vipc_client_main.height, True)
+        m.warmup()
+        big_model = m
+      except Exception:
+        cloudlog.exception("big model load failed")
+    loader = threading.Thread(target=load_big, daemon=True)
+    loader.start()
+    loader.join(BIG_MODEL_TIMEOUT)
+    model = big_model
+    params.put_bool("UsbGpuActive", model is not None)
+
+  small_model = ModelState(vipc_client_main.width, vipc_client_main.height, False) if model is None or USBGPU else None
+  if model is None:
+    model = small_model
+  params.put_bool("UsbGpuLoading", False)
   cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
 
   # messaging
-  pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"])
+  pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry"])
   sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"])
 
   publish_state = PublishState()
@@ -252,9 +274,7 @@ def main(demo=False):
     is_rhd = sm["driverMonitoringState"].isRHD
     frame_id = sm["roadCameraState"].frameId
     v_ego = max(sm["carState"].vEgo, 0.)
-    if sm.frame % 60 == 0:
+    lat_delay = sm["liveDelay"].lateralDelay + LAT_SMOOTH_SECONDS
-      model.lat_delay = get_lat_delay(params, sm["liveDelay"].lateralDelay)
-    lat_delay = model.lat_delay + LAT_SMOOTH_SECONDS
     if sm.updated["liveCalibration"] and sm.seen['roadCameraState'] and sm.seen['deviceState']:
       device_from_calib_euler = np.array(sm["liveCalibration"].rpyCalib, dtype=np.float32)
       dc = DEVICE_CAMERAS[(str(sm['deviceState'].deviceType), str(sm['roadCameraState'].sensor))]
@@ -293,7 +313,17 @@ def main(demo=False):
     }
 
     mt1 = time.perf_counter()
-    model_output = model.run(bufs, transforms, inputs)
+    try:
+      model_output = model.run(bufs, transforms, inputs)
+    except Exception:
+      if not params.get_bool("UsbGpuActive"):
+        raise
+      # fallback to small model
+      cloudlog.exception("big model failed, fall back to small")
+      params.put_bool("UsbGpuActive", False)
+      model = small_model
+      run_count = 0
+      model_output = None
     mt2 = time.perf_counter()
     model_execution_time = mt2 - mt1
 
@@ -301,7 +331,6 @@ def main(demo=False):
       modelv2_send = messaging.new_message('modelV2')
       drivingdata_send = messaging.new_message('drivingModelData')
       posenet_send = messaging.new_message('cameraOdometry')
-      mdv2sp_send = messaging.new_message('modelDataV2SP')
 
       action = get_action_from_model(model_output, prev_action, lat_action_t, long_action_t, v_ego)
       prev_action = action
@@ -316,14 +345,12 @@ def main(demo=False):
       DH.update(sm['carState'], sm['carControl'].latActive, lane_change_prob)
       modelv2_send.modelV2.meta.laneChangeState = DH.lane_change_state
       modelv2_send.modelV2.meta.laneChangeDirection = DH.lane_change_direction
-      mdv2sp_send.modelDataV2SP.laneTurnDirection = DH.lane_turn_direction
 
       fill_driving_model_data(drivingdata_send, modelv2_send)
       fill_pose_msg(posenet_send, model_output, meta_main.frame_id, vipc_dropped_frames, meta_main.timestamp_eof, live_calib_seen)
       pm.send('modelV2', modelv2_send)
       pm.send('drivingModelData', drivingdata_send)
       pm.send('cameraOdometry', posenet_send)
-      pm.send('modelDataV2SP', mdv2sp_send)
     last_vipc_frame_id = meta_main.frame_id
 
 

openpilot/selfdrive/selfdrived/selfdrived.py

@@ -30,6 +30,7 @@ from openpilot.sunnypilot import get_sanitize_int_param
 from openpilot.sunnypilot.selfdrive.car.car_specific import CarSpecificEventsSP
 from openpilot.sunnypilot.selfdrive.car.cruise_helpers import CruiseHelper
 from openpilot.sunnypilot.selfdrive.car.intelligent_cruise_button_management.controller import IntelligentCruiseButtonManagement
+from openpilot.sunnypilot.selfdrive.selfdrived.button_state_tracker import ButtonStateTracker
 from openpilot.sunnypilot.selfdrive.selfdrived.events import EventsSP
 
 REPLAY = "REPLAY" in os.environ
@@ -177,6 +178,7 @@ class SelfdriveD(CruiseHelper):
     self.car_events_sp = CarSpecificEventsSP(self.CP, self.CP_SP)
 
     CruiseHelper.__init__(self, self.CP)
+    self.button_state_tracker = ButtonStateTracker()
 
   def update_events(self, CS):
     """Compute onroadEvents from carState"""
@@ -597,6 +599,8 @@ class SelfdriveD(CruiseHelper):
     icbm.sendButton = self.icbm.cruise_button
     icbm.vTarget = self.icbm.v_target
 
+    self.button_state_tracker.publish(ss_sp)
+
     self.pm.send('selfdriveStateSP', ss_sp_msg)
 
     # onroadEventsSP - logged every second or on change
@@ -616,6 +620,7 @@ class SelfdriveD(CruiseHelper):
       self.mads.update(CS)
     self.update_alerts(CS)
 
+    self.button_state_tracker.update(CS)
     self.publish_selfdriveState(CS)
 
     self.CS_prev = CS

openpilot/selfdrive/ui/sunnypilot/layouts/settings/steering_sub_layouts/mads_settings.py

@@ -7,7 +7,7 @@ See the LICENSE.md file in the root directory for more details.
 from collections.abc import Callable
 import pyray as rl
 
-from opendbc.sunnypilot.car.tesla.values import TeslaFlagsSP
+from opendbc.sunnypilot.car.tesla.values import MadsScreenButtonType, TeslaFlagsSP
 from openpilot.selfdrive.ui.ui_state import ui_state
 from openpilot.sunnypilot.mads.helpers import MadsSteeringModeOnBrake
 from openpilot.system.ui.lib.multilang import tr, tr_noop
@@ -96,7 +96,10 @@ class MadsSettingsLayout(Widget):
     if brand == "rivian":
       return True
     elif brand == "tesla":
-      return not (ui_state.CP_SP is not None and ui_state.CP_SP.flags & TeslaFlagsSP.HAS_VEHICLE_BUS)
+      if ui_state.CP_SP is None or not ui_state.CP_SP.flags & TeslaFlagsSP.HAS_VEHICLE_BUS:
+        return True
+      screen_button = int(ui_state.params.get("TeslaMadsScreenButton", return_default=True))
+      return screen_button == MadsScreenButtonType.OFF
     return False
 
   def _update_steering_mode_description(self, button_index: int):

openpilot/selfdrive/ui/sunnypilot/layouts/settings/vehicle/brands/tesla.py

@@ -4,10 +4,11 @@ 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.
 """
+from opendbc.sunnypilot.car.tesla.values import TeslaFlagsSP
 from openpilot.selfdrive.ui.sunnypilot.layouts.settings.vehicle.brands.base import BrandSettings
 from openpilot.selfdrive.ui.ui_state import ui_state
 from openpilot.system.ui.lib.multilang import tr
-from openpilot.system.ui.sunnypilot.widgets.list_view import toggle_item_sp
+from openpilot.system.ui.sunnypilot.widgets.list_view import multiple_button_item_sp, toggle_item_sp
 
 COOP_STEERING_MIN_KMH = 23
 OEM_STEERING_MIN_KMH = 48
@@ -18,7 +19,14 @@ class TeslaSettings(BrandSettings):
   def __init__(self):
     super().__init__()
     self.coop_steering_toggle = toggle_item_sp(tr("Cooperative Steering (Beta)"), "", param="TeslaCoopSteering")
-    self.items = [self.coop_steering_toggle]
+    self.mads_screen_button = multiple_button_item_sp(
+      title=lambda: tr("MADS Screen Activation"),
+      description="",
+      buttons=[lambda: tr("Off"), lambda: tr("3-Finger"), lambda: tr("4-Finger"), lambda: tr("5-Finger")],
+      param="TeslaMadsScreenButton",
+      inline=False,
+    )
+    self.items = [self.coop_steering_toggle, self.mads_screen_button]
 
   def update_settings(self):
     is_metric = ui_state.is_metric
@@ -41,3 +49,18 @@ class TeslaSettings(BrandSettings):
 
     self.coop_steering_toggle.set_description(coop_steering_desc)
     self.coop_steering_toggle.action_item.set_enabled(ui_state.is_offroad())
+
+    has_vehicle_bus = ui_state.CP_SP is not None and bool(ui_state.CP_SP.flags & TeslaFlagsSP.HAS_VEHICLE_BUS)
+    self.mads_screen_button.set_visible(has_vehicle_bus)
+
+    mads_screen_button_desc = (
+      f"{tr('Use a multi-finger press on the infotainment screen to toggle MADS.')} " +
+      f"{tr('This allows the use of full MADS functionality when enabled.')}<br><br>" +
+      f"{tr('Selecting a higher finger count may reduce accidental activations.')}<br><br>" +
+      f"<b>{tr('Note: Setting this to Off will reset your MADS settings to default.')}</b>"
+    )
+    if not ui_state.is_offroad():
+      mads_screen_button_disabled_msg = tr("Enable \"Always Offroad\" in Device panel, or turn vehicle off to change.")
+      mads_screen_button_desc = f"<b>{mads_screen_button_disabled_msg}</b><br><br>{mads_screen_button_desc}"
+    self.mads_screen_button.set_description(mads_screen_button_desc)
+    self.mads_screen_button.action_item.set_enabled(ui_state.is_offroad())

openpilot/sunnypilot/mads/helpers.py

@@ -9,7 +9,7 @@ from openpilot.common.params import Params
 from opendbc.car import structs
 from opendbc.safety import ALTERNATIVE_EXPERIENCE
 from opendbc.sunnypilot.car.hyundai.values import HyundaiFlagsSP, HyundaiSafetyFlagsSP
-from opendbc.sunnypilot.car.tesla.values import TeslaFlagsSP
+from opendbc.sunnypilot.car.tesla.values import MadsScreenButtonType, TeslaFlagsSP
 
 
 MADS_NO_ACC_MAIN_BUTTON = ("rivian", "tesla")
@@ -21,17 +21,20 @@ class MadsSteeringModeOnBrake:
   DISENGAGE = 2
 
 
-def get_mads_limited_brands(CP: structs.CarParams, CP_SP: structs.CarParamsSP) -> bool:
+def get_mads_limited_brands(CP: structs.CarParams, CP_SP: structs.CarParamsSP, params: Params) -> bool:
   if CP.brand == 'rivian':
     return True
   if CP.brand == 'tesla':
-    return not CP_SP.flags & TeslaFlagsSP.HAS_VEHICLE_BUS
+    if not CP_SP.flags & TeslaFlagsSP.HAS_VEHICLE_BUS:
+      return True
+    screen_button = int(params.get("TeslaMadsScreenButton", return_default=True))
+    return screen_button == MadsScreenButtonType.OFF
 
   return False
 
 
 def read_steering_mode_param(CP: structs.CarParams, CP_SP: structs.CarParamsSP, params: Params):
-  if get_mads_limited_brands(CP, CP_SP):
+  if get_mads_limited_brands(CP, CP_SP, params):
     return MadsSteeringModeOnBrake.DISENGAGE
 
   return params.get("MadsSteeringMode", return_default=True)
@@ -63,7 +66,7 @@ def set_car_specific_params(CP: structs.CarParams, CP_SP: structs.CarParamsSP, p
   # MADS is currently partially supported for these platforms due to lack of consistent states to engage controls
   # Only MadsSteeringModeOnBrake.DISENGAGE is supported for these platforms
   # TODO-SP: To enable MADS full support for Rivian and most Tesla, identify consistent signals for MADS toggling
-  mads_partial_support = get_mads_limited_brands(CP, CP_SP)
+  mads_partial_support = get_mads_limited_brands(CP, CP_SP, params)
   if mads_partial_support:
     params.put("MadsSteeringMode", 2, block=True)
     params.put_bool("MadsUnifiedEngagementMode", True, block=True)

openpilot/sunnypilot/mads/tests/test_mads_steering_mode.py

@@ -13,7 +13,7 @@ from openpilot.selfdrive.selfdrived.events import Events
 from openpilot.sunnypilot.selfdrive.selfdrived.events import EventsSP
 from openpilot.sunnypilot.mads.helpers import MadsSteeringModeOnBrake, read_steering_mode_param
 from openpilot.sunnypilot.mads.mads import ModularAssistiveDrivingSystem
-from opendbc.sunnypilot.car.tesla.values import TeslaFlagsSP
+from opendbc.sunnypilot.car.tesla.values import MadsScreenButtonType, TeslaFlagsSP
 
 State = custom.ModularAssistiveDrivingSystem.ModularAssistiveDrivingSystemState
 EventName = log.OnroadEvent.EventName
@@ -38,6 +38,12 @@ def make_panda_state(mocker, controls_allowed_lateral=True):
   return ps
 
 
+def make_params_mock(mocker, values):
+  params = mocker.MagicMock()
+  params.get = mocker.MagicMock(side_effect=lambda k, **kwargs: values[k])
+  return params
+
+
 def make_mads(mocker, steering_mode):
   sd = mocker.MagicMock()
   sd.CP = structs.CarParams()
@@ -223,15 +229,27 @@ class TestBrandSteeringModeRestrictions:
     params = mocker.MagicMock()
     assert read_steering_mode_param(CP, CP_SP, params) == MadsSteeringModeOnBrake.DISENGAGE
 
-  def test_tesla_with_vehicle_bus_uses_param(self, mocker):
+  @pytest.mark.parametrize("screen_button", [MadsScreenButtonType.THREE_FINGER,
+                                             MadsScreenButtonType.FOUR_FINGER,
+                                             MadsScreenButtonType.FIVE_FINGER])
+  def test_tesla_with_vehicle_bus_uses_param(self, mocker, screen_button):
     CP = structs.CarParams()
     CP.brand = "tesla"
     CP_SP = structs.CarParamsSP()
     CP_SP.flags = TeslaFlagsSP.HAS_VEHICLE_BUS
-    params = mocker.MagicMock()
+    params = make_params_mock(mocker, {"TeslaMadsScreenButton": screen_button,
-    params.get = mocker.MagicMock(return_value=MadsSteeringModeOnBrake.REMAIN_ACTIVE)
+                                       "MadsSteeringMode": MadsSteeringModeOnBrake.REMAIN_ACTIVE})
     assert read_steering_mode_param(CP, CP_SP, params) == MadsSteeringModeOnBrake.REMAIN_ACTIVE
 
+  def test_tesla_with_vehicle_bus_screen_button_off_forced_to_disengage(self, mocker):
+    CP = structs.CarParams()
+    CP.brand = "tesla"
+    CP_SP = structs.CarParamsSP()
+    CP_SP.flags = TeslaFlagsSP.HAS_VEHICLE_BUS
+    params = make_params_mock(mocker, {"TeslaMadsScreenButton": MadsScreenButtonType.OFF,
+                                       "MadsSteeringMode": MadsSteeringModeOnBrake.REMAIN_ACTIVE})
+    assert read_steering_mode_param(CP, CP_SP, params) == MadsSteeringModeOnBrake.DISENGAGE
+
   @pytest.mark.parametrize("brand", ["hyundai", "toyota", "honda", "gm"])
   def test_other_brands_use_param(self, mocker, brand):
     CP = structs.CarParams()

openpilot/sunnypilot/modeld_v2/SConscript

@@ -1,9 +1,12 @@
 import os
 import glob
+import sys
+import subprocess
 
 from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
 from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE
 from openpilot.common.hardware import HARDWARE, PC
+from openpilot.selfdrive.modeld.helpers import usbgpu_present
 
 Import('env', 'arch', 'release')
 lenv = env.Clone()
@@ -22,14 +25,21 @@ def get_camera_configs():
 
 CAMERA_CONFIGS = get_camera_configs()
 
-tg_flags = {
+def probe_devices():
-  'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0',
+  return set(subprocess.run(
-  'Darwin': f'DEV=CPU HOME={os.path.expanduser("~")}',
+    [sys.executable, '-c', 'from tinygrad import Device\nprint("\\n".join(Device.get_available_devices()))'],
-}.get(arch, 'DEV=CPU:LLVM')
+    capture_output=True, text=True, check=True).stdout.strip().splitlines())
 
-image_flag = {
+available = probe_devices()
-  'larch64': 'IMAGE=2',
+if 'CUDA' in available:
-}.get(arch, 'IMAGE=0')
+  tg_backend = 'CUDA'
+  tg_flags = f'DEV={tg_backend}'
+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'
+else:
+  tg_backend = 'CPU'
+  tg_flags = f'DEV=CPU HOME={os.path.expanduser("~")}' if arch == 'Darwin' else 'DEV=CPU:LLVM'
 
 model_w, model_h = MEDMODEL_INPUT_SIZE
 from openpilot.selfdrive.modeld.constants import ModelConstants
@@ -41,17 +51,30 @@ compile_modeld_script = File("compile_modeld.py").abspath
 upstream_compile_script = File(Dir("#openpilot/selfdrive/modeld").File("compile_modeld.py").abspath)
 script_deps = [File("compile_modeld.py"), upstream_compile_script]
 
+USBGPU = usbgpu_present()
+if USBGPU:
+  usbgpu_tg_flags = f'DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV={tg_backend} FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0'
+  usbgpu_lock = File("models/.usb_gpu.lock").abspath
+
 def compile_combined(model_type, onnx_args, output_name):
-  output_pkl = File(f"models/{output_name}").abspath
+  for usbgpu in ([False, True] if USBGPU else [False]):
-  cmd = (f'{pythonpath_string} {tg_flags} {image_flag} python3 {compile_modeld_script} '
+    prefix = 'big_' if usbgpu else ('big_' if os.getenv('BIG_INTO_SMALL') else '')
-         f'--model-type {model_type} '
+    final_output_name = prefix + output_name
-         f'--model-size {model_w}x{model_h} '
+    output_pkl = File(f"models/{final_output_name}").abspath
-         f'--camera-resolutions {camera_res_args} '
+
-         f'{onnx_args} '
+    active_tg_flags = usbgpu_tg_flags if usbgpu else tg_flags
-         f'--frame-skip {frame_skip} '
+
-         f'--output {output_pkl}')
+    cmd = (f'{pythonpath_string} {active_tg_flags} python3 {compile_modeld_script} '
-  onnx_files = [f for f in onnx_args.split() if f.endswith('.onnx')]
+           f'--model-type {model_type} '
-  return lenv.Command(output_pkl, tinygrad_files + script_deps + [File(f) for f in onnx_files if os.path.isfile(f)], cmd)
+           f'--model-size {model_w}x{model_h} '
+           f'--camera-resolutions {camera_res_args} '
+           f'{onnx_args} '
+           f'--frame-skip {frame_skip} '
+           f'--output {output_pkl}')
+    onnx_files = [f for f in onnx_args.split() if f.endswith('.onnx')]
+    node = lenv.Command(output_pkl, tinygrad_files + script_deps + [File(f) for f in onnx_files if os.path.isfile(f)], cmd)
+    if usbgpu:
+      lenv.SideEffect(usbgpu_lock, node)
 
 # Vision + Policy (stock default model)
 vision_onnx = File("models/driving_vision.onnx").abspath
@@ -82,18 +105,3 @@ if os.path.isfile(supercombo_onnx):
   compile_combined('supercombo',
                    f'--supercombo-onnx {supercombo_onnx}',
                    'driving_combined_supercombo_tinygrad.pkl')
-
-if PC:
-  inputs = tinygrad_files + [File(Dir("#openpilot/sunnypilot/modeld_v2").File("install_models_pc.py").abspath)]
-  outputs = []
-  model_dir = Dir("models").abspath
-  cmd = f'python3 {Dir("#openpilot/sunnypilot/modeld_v2").abspath}/install_models_pc.py {model_dir}'
-
-  for model_name in ['supercombo', 'driving_vision', 'driving_off_policy', 'driving_on_policy', 'driving_policy']:
-    if File(f"models/{model_name}.onnx").exists():
-        inputs.append(File(f"models/{model_name}.onnx"))
-        inputs.append(File(f"models/{model_name}_tinygrad.pkl"))
-        outputs.append(File(f"models/{model_name}_metadata.pkl"))
-  if outputs:
-    lenv.Command(outputs, inputs, cmd)
-

openpilot/sunnypilot/modeld_v2/compile_modeld.py

@@ -8,10 +8,10 @@ See the LICENSE.md file in the root directory for more details.
 
 import argparse
 import os
-import pickle
+import tempfile
-import time
 from collections import defaultdict
 from functools import partial
+from openpilot.selfdrive.modeld.helpers import dump_oob, load_oob
 import numpy as np
 os.environ['GMMU'] = '0'
 
@@ -57,7 +57,26 @@ def derive_frame_skip(vision_input_shapes: dict, policy_input_shapes: dict) -> i
   return 1 if not features_buffer or features_buffer[1] >= 99 else 4
 
 
-def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+def get_policy_npy_shapes(input_shapes: dict, is_supercombo: bool = False) -> tuple[dict, list[int]]:
+  desire_key = _detect_desire_key(input_shapes)
+  shapes = {}
+  if desire_key:
+    shapes['desire'] = (input_shapes[desire_key][2],)
+
+  if is_supercombo and 'features_buffer' in input_shapes:
+    fb = input_shapes['features_buffer']
+    shapes['prev_feat'] = (fb[0], fb[2])
+
+  for key, shape in input_shapes.items():
+    if key not in (desire_key, 'features_buffer') and 'img' not in key:
+      shapes[key] = tuple(shape)
+
+  sizes = [int(np.prod(size)) for size in shapes.values()]
+  return shapes, sizes
+
+
+def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT,
+                            is_supercombo: bool = False) -> tuple[dict, dict]:
   road_key, _ = _detect_vision_keys(input_shapes)
   if not road_key:
     raise ValueError("Vision road key missing from input shapes.")
@@ -74,36 +93,43 @@ def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = D
   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)
   }
 
-  for key, shape in input_shapes.items():
+  shapes, sizes = get_policy_npy_shapes(input_shapes, is_supercombo=is_supercombo)
-    if key not in npy_arrays and 'img' not in key and key not in ('features_buffer', desire_key):
+  packed_npy_inputs = np.zeros(sum(sizes), dtype=np.float32)
-      npy_arrays[key] = np.zeros(shape, dtype=np.float32)
+
+  split_indices = np.cumsum(sizes[:-1]) if len(sizes) > 1 else []
+  split_views = np.split(packed_npy_inputs, split_indices) if len(sizes) > 0 else []
+  for (k, s), v in zip(shapes.items(), split_views, strict=True):
+    npy_arrays[k] = v.reshape(s)
 
   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]),
-                  dtype=np.float32), device=device).contiguous().realize()
+                  dtype=np.float32), device=device).contiguous().realize(),
+    'packed_npy_inputs': Tensor(packed_npy_inputs, device='NPY').realize(),
   }
 
   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()})
+  queues.update({key: Tensor(value, device='NPY').realize() for key, value in npy_arrays.items() if key in ('tfm', 'big_tfm')})
+
   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_split_input_queues(vision_input_shapes: dict, policy_input_shapes: dict,
-  return generate_queues_and_npy({**vision_input_shapes, **policy_input_shapes}, frame_skip, device)
+                            frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+  return generate_queues_and_npy({**vision_input_shapes, **policy_input_shapes}, frame_skip, device, is_supercombo=False)
 
 
-def make_supercombo_input_queues(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+def make_supercombo_input_queues(input_shapes: dict, frame_skip: int,
-  return generate_queues_and_npy(input_shapes, frame_skip, device)
+                                 device: str = Device.DEFAULT) -> tuple[dict, dict]:
+  return generate_queues_and_npy(input_shapes, frame_skip, device, is_supercombo=True)
 
 
 def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, model_size: tuple[int, int],
@@ -118,22 +144,17 @@ def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, mode
   if not desire_key or not road_key or not wide_key:
     raise ValueError("Missing required vision or desire keys in input shapes.")
 
-  extra_keys = [key for key in input_shapes if key not in (desire_key, 'features_buffer', 'traffic_convention') and 'img' not in key]
+  is_supercombo = vision_runner is None
+  npy_shapes, npy_sizes = get_policy_npy_shapes(input_shapes, is_supercombo=is_supercombo)
 
-  def runner(img_q, big_img_q, feat_q, frame, big_frame, tfm, big_tfm, **kwargs):
+  def runner(img_q, big_img_q, feat_q, packed_npy_inputs, 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)]
+    packed_npy_inputs_dev = packed_npy_inputs.to(Device.DEFAULT)
-    if traffic_convention is not None:
+    tfm_dev = tfm.to(Device.DEFAULT)
-      npys.append(traffic_convention.to(Device.DEFAULT))
+    big_tfm_dev = big_tfm.to(Device.DEFAULT)
 
-    extra_tensors = {key: kwargs[key].to(Device.DEFAULT) for key in extra_keys if key in kwargs}
+    Tensor.realize(packed_npy_inputs_dev, tfm_dev, big_tfm_dev)
-    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).realize()
     big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm_dev).unsqueeze(0), sample_skip_fn).realize()
@@ -141,22 +162,37 @@ def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, mode
     if prepare_only:
       return img, big_img
 
-    desire_buf = shift_and_sample(desire_q, desire_dev.reshape(1, 1, -1), sample_desire_fn).realize()
+    unpacked_tensors = [tensor.reshape(shape) for tensor, shape in zip(packed_npy_inputs_dev.split(npy_sizes), npy_shapes.values(), strict=True)]
-    inputs = {desire_key: desire_buf, **extra_tensors}
+    unpacked_dict = dict(zip(npy_shapes.keys(), unpacked_tensors, strict=True))
 
-    if traffic_conv_dev is not None:
+    desire_dev = unpacked_dict['desire']
-      inputs['traffic_convention'] = traffic_conv_dev
+    desire_buf = shift_and_sample(desire_q, desire_dev.reshape(1, 1, -1), sample_desire_fn).realize()
+
+    inputs = {desire_key: desire_buf}
+    for key, tensor_val in unpacked_dict.items():
+      if key not in ('desire', 'prev_feat'):
+        inputs[key] = tensor_val
+
+    if 'prev_feat' in unpacked_dict:
+      prev_feat_dev = unpacked_dict['prev_feat']
+      inputs['features_buffer'] = shift_and_sample(feat_q, prev_feat_dev.reshape(1, 1, -1), sample_skip_fn).realize()
 
     if vision_runner:
       vision_out_cast = next(iter(vision_runner({road_key: img, wide_key: big_img}).values())).cast('float32').realize()
-      new_feat = vision_out_cast[:, features_slice].reshape(1, -1).unsqueeze(0)
+      if 'features_buffer' not in inputs:
-      inputs['features_buffer'] = shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
+        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').realize() 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)})
+
+    inputs.update({road_key: img, wide_key: big_img})
+    if 'features_buffer' not in inputs:
+      inputs['features_buffer'] = sample_skip_fn(feat_q)
+
     policy_out = next(iter(policy_runners[0](inputs).values())).cast('float32').realize()
-    new_feat = policy_out[:, features_slice].reshape(1, -1).unsqueeze(0)
+    if 'features_buffer' not in inputs and features_slice is not None:
-    shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
+      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
@@ -172,27 +208,34 @@ def compile_and_warmup(nv12: NV12Frame, model_size: tuple[int, int], prepare_onl
     raise ValueError("Could not find vision, model, or policy metadata.")
 
   features_slice = feat_meta['output_slices']['hidden_state']
-  WARP_DEV = 'CPU' if "USBGPU" in os.environ else Device.DEFAULT
+  WARP_DEV = os.getenv('WARP_DEV', Device.DEFAULT)
 
+  is_supercombo = vision_runner is None
   run_func = create_jit_runner(vision_runner, policy_runners, nv12, model_size, features_slice, frame_skip, all_shapes, prepare_only)
   run_jit = TinyJit(run_func, prune=True)
-  queues, npy_arrays = generate_queues_and_npy(all_shapes, frame_skip, Device.DEFAULT)
+  def run_once(seed):
-
+    queues, npy = generate_queues_and_npy(all_shapes, frame_skip, Device.DEFAULT, is_supercombo=is_supercombo)
-  for i in range(3):
+    rng = np.random.default_rng(seed)
-    np.random.seed(42 + i)
     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():
+    for value in npy.values():
-      arr[:] = np.random.randn(*arr.shape).astype(arr.dtype)
+      value[:] = rng.standard_normal(value.shape).astype(value.dtype)
-
     Device.default.synchronize()
-    start_time = time.perf_counter()
+    outs = run_jit(**queues, frame=frame, big_frame=big_frame)
-    run_jit(**queues, frame=frame, big_frame=big_frame)
-    mid_time = time.perf_counter()
     Device.default.synchronize()
-    print(f"  [{i + 1}/3] enqueue {(mid_time - start_time) * 1e3:6.2f} ms -- total {(time.perf_counter() - start_time) * 1e3:6.2f} ms")
+    return [np.copy(value.numpy()) for value in (outs if isinstance(outs, tuple) else [outs])] if outs is not None else []
 
-  return pickle.loads(pickle.dumps(run_jit)) if not prepare_only else run_jit
+  for i in range(3):
+    run_once(42 + i)
+
+  if not prepare_only:
+    baseline = run_once(42)
+    with tempfile.TemporaryFile(dir=".") as f:
+      dump_oob(run_jit, f)
+      f.seek(0)
+      run_jit = load_oob(f)
+    assert all(np.array_equal(baseline, deserialized) for baseline, deserialized in zip(baseline, run_once(42), strict=True)), "OOB pickling regression"
+  return run_jit
 
 
 def _parse_size(size_str: str) -> tuple[int, int]:
@@ -289,7 +332,7 @@ if __name__ == "__main__":
                         vision_runner, policy_runners, output_data['metadata']))
 
   with open(args.output, "wb") as file:
-    pickle.dump(output_data, file)
+    dump_oob(output_data, file)
 
   pkl_size = os.path.getsize(args.output)
   print(f"Saved combined JIT to {args.output} ({pkl_size / 1e6:.2f} MB)")

openpilot/sunnypilot/modeld_v2/install_models_pc.py

@@ -1,75 +0,0 @@
-#!/usr/bin/env python3
-import sys
-import shutil
-import pickle
-import codecs
-from pathlib import Path
-
-from openpilot.common.hardware.hw import Paths
-from openpilot.sunnypilot.modeld_v2.get_model_metadata import MetadataOnnxPBParser, get_name_and_shape, get_metadata_value_by_name
-
-
-def generate_metadata_pkl(model_path, output_path):
-  try:
-    model = MetadataOnnxPBParser(model_path).parse()
-    output_slices = get_metadata_value_by_name(model, 'output_slices')
-    if not output_slices:
-      return False
-    metadata = {
-      'model_checkpoint': get_metadata_value_by_name(model, 'model_checkpoint'),
-      'output_slices': pickle.loads(codecs.decode(output_slices.encode(), "base64")),
-      'input_shapes': dict(get_name_and_shape(x) for x in model["graph"]["input"]),
-      'output_shapes': dict(get_name_and_shape(x) for x in model["graph"]["output"]),
-    }
-    with open(output_path, 'wb') as f:
-      pickle.dump(metadata, f)
-    return True
-  except Exception:
-    return False
-
-
-def install_models(model_dir):
-  model_dir = Path(model_dir)
-  models = ["driving_off_policy", "driving_on_policy", "driving_vision"]
-  found_models = []
-
-  for model in models:
-    if (model_dir / f"{model}.onnx").exists():
-      found_models.append(model)
-
-  if not found_models:
-    return
-
-  try:
-    custom_name = input(f"Found models ({', '.join(found_models)}). Enter model short name (e.g. wmiv4): ").strip()
-  except EOFError:
-    return
-
-  if not custom_name:
-    print("No name provided, skipping installation.")
-    return
-
-  dest_dir = Path(Paths.model_root())
-  dest_dir.mkdir(parents=True, exist_ok=True)
-
-  for model in found_models:
-    onnx_path = model_dir / f"{model}.onnx"
-    tinygrad_pkl = model_dir / f"{model}_tinygrad.pkl"
-    metadata_pkl = model_dir / f"{model}_metadata.pkl"
-
-    if not metadata_pkl.exists():
-      generate_metadata_pkl(onnx_path, metadata_pkl)
-
-    dest_tinygrad = dest_dir / f"{model}_{custom_name}_tinygrad.pkl"
-    dest_metadata = dest_dir / f"{model}_{custom_name}_metadata.pkl"
-
-    if tinygrad_pkl.exists():
-      shutil.move(str(tinygrad_pkl), str(dest_tinygrad))
-    if metadata_pkl.exists():
-      shutil.move(str(metadata_pkl), str(dest_metadata))
-
-if __name__ == "__main__":
-  if len(sys.argv) < 2:
-    print("Usage: install_models_pc.py <model_dir>")
-    sys.exit(1)
-  install_models(sys.argv[1])

openpilot/sunnypilot/modeld_v2/modeld.py

@@ -10,11 +10,14 @@ 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
+
+from openpilot.selfdrive.modeld.helpers import usbgpu_present, load_oob
+from openpilot.selfdrive.modeld.usbgpu_link import wait_usbgpu_link
+
+USBGPU = usbgpu_present()
 if USBGPU:
   os.environ['DEV'] = 'AMD'
   os.environ['AMD_IFACE'] = 'USB'
-import pickle
 import time
 import numpy as np
 import openpilot.cereal.messaging as messaging
@@ -24,6 +27,11 @@ from setproctitle import setproctitle
 from openpilot.cereal.messaging import PubMaster, SubMaster
 from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
 from opendbc.car.car_helpers import get_demo_car_params
+
+from tinygrad.tensor import Tensor
+from tinygrad.device import Device
+
+from openpilot.common.file_chunker import open_file_chunked
 from openpilot.common.swaglog import cloudlog
 from openpilot.common.params import Params
 from openpilot.common.filter_simple import FirstOrderFilter
@@ -31,6 +39,7 @@ from openpilot.common.realtime import config_realtime_process, DT_MDL
 from openpilot.common.transformations.camera import DEVICE_CAMERAS
 from openpilot.common.transformations.model import get_warp_matrix
 from openpilot.system import sentry
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value
 
@@ -38,6 +47,7 @@ from openpilot.sunnypilot.modeld_v2.fill_model_msg import fill_model_msg, fill_p
 from openpilot.sunnypilot.modeld_v2.constants import Plan
 from openpilot.sunnypilot.modeld_v2.meta_helper import load_meta_constants
 from openpilot.sunnypilot.modeld_v2.camera_offset_helper import CameraOffsetHelper
+from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues
 
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
@@ -100,31 +110,29 @@ class ModelState(ModelStateBase):
     self._init_combined(pkl_path, cam_w, cam_h, model_bundle)
 
   def _init_combined(self, pkl_path, cam_w, cam_h, bundle):
-    from tinygrad.tensor import Tensor
-    from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-    from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues
-    from tinygrad.device import Device
-
-    from openpilot.common.file_chunker import open_file_chunked
-
     cloudlog.warning(f"loading combined pkl: {pkl_path}")
-    jits = pickle.load(open_file_chunked(pkl_path))
+    jits = load_oob(open_file_chunked(pkl_path))
 
     self.DEV = Device.DEFAULT
-    self.WARP_DEV = 'CPU' if USBGPU else self.DEV
+    self.WARP_DEV = ('QCOM' if TICI else 'CPU') if USBGPU else self.DEV
     self.QUEUE_DEV = self.DEV
 
     metadata = jits['metadata']
+
+    self._run_policy = jits[(cam_w, cam_h)]['run_policy']
+    self._warp_enqueue = jits[(cam_w, cam_h)]['warp_enqueue']
+
     if 'model' in metadata:
       model_metadata = metadata['model']
       self.vision_output_slices = model_metadata['output_slices']
       self.policy_output_slices = {}
       self._policy_slices_list = []
       self._combined_model_type = 'supercombo'
-      self._vision_input_names = [k for k in model_metadata['input_shapes'] if 'img' in k]
+      self._vision_input_names = [key for key in model_metadata['input_shapes'] if 'img' in key]
       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.QUEUE_DEV)
     else:
       vision_metadata = metadata['vision']
       policy_keys = [k for k in metadata if k != 'vision']
@@ -142,16 +150,13 @@ 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.QUEUE_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
-    self.parser = SplitParser() if self._combined_model_type != 'supercombo' else CombinedParser()
-
     is_20hz = bundle.is20hz if bundle else self._combined_model_type in ('split', 'multi_policy')
     if is_20hz:
       from openpilot.sunnypilot.models.split_model_constants import SplitModelConstants
@@ -160,20 +165,45 @@ class ModelState(ModelStateBase):
       from openpilot.sunnypilot.modeld_v2.constants import ModelConstants
       self.constants = ModelConstants()
 
+    if self._combined_model_type != 'supercombo':
+      from openpilot.sunnypilot.modeld_v2.parse_model_outputs_split import Parser as SplitParser
+      self.parser = SplitParser()
+    else:
+      from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser as CombinedParser
+      self.parser = CombinedParser()
+
     self.prev_desire = np.zeros(self.constants.DESIRE_LEN, dtype=np.float32)
     self.full_frames: dict = {}
     self._blob_cache: dict = {}
     nv12_info = get_nv12_info(cam_w, cam_h)
     self.frame_buf_params = dict.fromkeys(self._vision_input_names, nv12_info)
 
-    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]
     self._warp_enqueue(
       **self.input_queues,
       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.WARP_DEV).contiguous().realize())
 
+    if USBGPU:
+      self.warmup()
+
+  def warmup(self) -> None:
+    dummy_frames = {k: np.zeros(self.frame_buf_params[k][3], dtype=np.uint8) for k in self._vision_input_names}
+    transforms = {k: np.eye(3, dtype=np.float32) for k in [self._road_key, self._wide_key] if k}
+
+    dummy_inputs = {}
+    for k, v in self.numpy_inputs.items():
+      if k not in ['tfm', 'big_tfm', 'prev_feat']:
+        dummy_inputs[k] = np.zeros(v.shape, dtype=v.dtype)
+
+    self.run(dummy_frames, transforms, dummy_inputs, prepare_only=False)
+
+    for v in self.numpy_inputs.values():
+      v[:] = 0
+    self.prev_desire[:] = 0
+    self.full_frames.clear()
+    self._blob_cache.clear()
+
 
   @property
   def mlsim(self) -> bool:
@@ -189,8 +219,6 @@ class ModelState(ModelStateBase):
 
   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:
-    from tinygrad.tensor import Tensor
-
     for key in bufs.keys():
       ptr = np.frombuffer(bufs[key].data, dtype=np.uint8).ctypes.data
       yuv_size = self.frame_buf_params[key][3]
@@ -222,11 +250,16 @@ class ModelState(ModelStateBase):
       model_output = raw_outputs.numpy().flatten()
       sliced = {k: model_output[np.newaxis, v] for k, v in self.vision_output_slices.items()}
       outputs = self.parser.parse_outputs(sliced)
+      if 'prev_feat' in self.numpy_inputs:
+        self.numpy_inputs['prev_feat'][:] = model_output[self.vision_output_slices['hidden_state']]
     else:
       vision_output = raw_outputs[0].numpy().flatten()
       vision_sliced = {k: vision_output[np.newaxis, v] for k, v in self.vision_output_slices.items()}
       outputs = self.parser.parse_vision_outputs(vision_sliced)
 
+      if 'prev_feat' in self.numpy_inputs and 'hidden_state' in self.vision_output_slices:
+        self.numpy_inputs['prev_feat'][:] = vision_output[self.vision_output_slices['hidden_state']]
+
       for i, policy_slices in enumerate(self._policy_slices_list):
         policy_output = raw_outputs[i + 1].numpy().flatten()
         policy_sliced = {k: policy_output[np.newaxis, v] for k, v in policy_slices.items()}
@@ -255,7 +288,6 @@ class ModelState(ModelStateBase):
       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,
                                                        action_t=long_action_t)
-      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)
@@ -265,6 +297,8 @@ class ModelState(ModelStateBase):
       desired_curvature = model_output['action'][0, 0] / (max(1.0, v_ego))**2
       should_stop = (v_ego < 0.3 and desired_accel < 0.1)
 
+    desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, self.LONG_SMOOTH_SECONDS)
+
     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)
@@ -282,6 +316,9 @@ def main(demo=False):
   setproctitle(PROCESS_NAME)
   config_realtime_process(7, 54)
 
+  if USBGPU:
+    wait_usbgpu_link()
+
   # visionipc clients
   while True:
     available_streams = VisionIpcClient.available_streams("camerad", block=False)
@@ -316,6 +353,9 @@ def main(demo=False):
   publish_state = PublishState()
   params = Params()
 
+  params.put_bool("UsbGpuPresent", USBGPU)
+  params.put_bool("UsbGpuCompiled", USBGPU)
+
   # setup filter to track dropped frames
   frame_dropped_filter = FirstOrderFilter(0., 10., 1. / model.constants.MODEL_FREQ)
   frame_id = 0

openpilot/sunnypilot/modeld_v2/parse_model_outputs.py

@@ -1,13 +1,16 @@
 import numpy as np
 from openpilot.sunnypilot.modeld_v2.constants import ModelConstants
 
+
 def safe_exp(x, out=None):
   # -11 is around 10**14, more causes float16 overflow
   return np.exp(np.clip(x, -np.inf, 11), out=out)
 
+
 def sigmoid(x):
   return 1. / (1. + safe_exp(-x))
 
+
 def softmax(x, axis=-1):
   x -= np.max(x, axis=axis, keepdims=True)
   if x.dtype == np.float32 or x.dtype == np.float64:
@@ -17,6 +20,19 @@ def softmax(x, axis=-1):
   x /= np.sum(x, axis=axis, keepdims=True)
   return x
 
+
+def _infer_mhp(slice_size: int, prod_out_shape: int, max_in_n: int = 16, max_out_n: int = 6) -> tuple[int, int]:
+  for out_n in range(max_out_n + 1):
+    per = 2 * prod_out_shape + out_n
+    if per <= 0:
+      continue
+    if slice_size % per == 0:
+      in_n = slice_size // per
+      if 1 <= in_n <= max_in_n:
+        return in_n, out_n
+  return 1, 0  # single hypothesis, no weights — matches a non-MDN output
+
+
 class Parser:
   def __init__(self, ignore_missing=False):
     self.ignore_missing = ignore_missing
@@ -40,17 +56,22 @@ class Parser:
     raw = outs[name]
     outs[name] = sigmoid(raw)
 
-  def parse_mdn(self, name, outs, in_N=0, out_N=1, out_shape=None):
+  def parse_mdn(self, name, outs, out_shape, in_N=0, out_N=0):
     if self.check_missing(outs, name):
       return
     raw = outs[name]
-    raw = raw.reshape((raw.shape[0], max(in_N, 1), -1))
+
+    if in_N == 0 and out_N == 0:
+      prod = int(np.prod(out_shape))
+      in_N, out_N = _infer_mhp(raw.shape[1], prod)
+
+    raw = raw.reshape((raw.shape[0], in_N, -1))
 
     n_values = (raw.shape[2] - out_N)//2
     pred_mu = raw[:,:,:n_values]
     pred_std = safe_exp(raw[:,:,n_values: 2*n_values])
 
-    if in_N > 1:
+    if in_N > 1 and out_N > 0:
       weights = np.zeros((raw.shape[0], in_N, out_N), dtype=raw.dtype)
       for i in range(out_N):
         weights[:,:,i - out_N] = softmax(raw[:,:,i - out_N], axis=-1)
@@ -61,7 +82,6 @@ class Parser:
           weights[fidx] = weights[fidx][idxs]
           pred_mu[fidx] = pred_mu[fidx][idxs]
           pred_std[fidx] = pred_std[fidx][idxs]
-      assert out_shape is not None
       full_shape = tuple([raw.shape[0], in_N] + list(out_shape))
       outs[name + '_weights'] = weights
       outs[name + '_hypotheses'] = pred_mu.reshape(full_shape)
@@ -74,37 +94,43 @@ class Parser:
           idxs = np.argsort(weights[fidx,:,hidx])[::-1]
           pred_mu_final[fidx, hidx] = pred_mu[fidx, idxs[0]]
           pred_std_final[fidx, hidx] = pred_std[fidx, idxs[0]]
+    elif in_N > 1 and out_N == 0:
+      # MHP without weights: keep every hypothesis intact, surface them as
+      # ``*_hypotheses`` and propagate the full multi-hypothesis tensor.
+      full_shape = tuple([raw.shape[0], in_N] + list(out_shape))
+      outs[name + '_hypotheses'] = pred_mu.reshape(full_shape)
+      outs[name + '_stds_hypotheses'] = pred_std.reshape(full_shape)
+      pred_mu_final = pred_mu
+      pred_std_final = pred_std
     else:
       pred_mu_final = pred_mu
       pred_std_final = pred_std
 
-    if out_N > 1:
+    if out_N > 1 or (in_N > 1 and out_N == 0):
-      assert out_shape is not None
+      n_selections = out_N if out_N > 1 else in_N
-      final_shape = tuple([raw.shape[0], out_N] + list(out_shape))
+      final_shape = tuple([raw.shape[0], n_selections] + list(out_shape))
     else:
-      assert out_shape is not None
       final_shape = tuple([raw.shape[0],] + list(out_shape))
     outs[name] = pred_mu_final.reshape(final_shape)
     outs[name + '_stds'] = pred_std_final.reshape(final_shape)
 
   def parse_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
-    self.parse_mdn('plan', outs, in_N=ModelConstants.PLAN_MHP_N, out_N=ModelConstants.PLAN_MHP_SELECTION,
+    # supercombo (4955 / 102) and newer variants (e.g. 990 / 144).
-                   out_shape=(ModelConstants.IDX_N,ModelConstants.PLAN_WIDTH))
+    self.parse_mdn('plan',   outs, out_shape=(ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH))
-    self.parse_mdn('lane_lines', outs, in_N=0, out_N=0, out_shape=(ModelConstants.NUM_LANE_LINES,ModelConstants.IDX_N,ModelConstants.LANE_LINES_WIDTH))
+    self.parse_mdn('lane_lines', outs, out_shape=(ModelConstants.NUM_LANE_LINES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
-    self.parse_mdn('road_edges', outs, in_N=0, out_N=0, out_shape=(ModelConstants.NUM_ROAD_EDGES,ModelConstants.IDX_N,ModelConstants.LANE_LINES_WIDTH))
+    self.parse_mdn('road_edges', outs, out_shape=(ModelConstants.NUM_ROAD_EDGES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
-    self.parse_mdn('pose', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+    self.parse_mdn('pose',         outs, out_shape=(ModelConstants.POSE_WIDTH,))
-    self.parse_mdn('road_transform', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+    self.parse_mdn('road_transform', outs, out_shape=(ModelConstants.POSE_WIDTH,))
     if 'sim_pose' in outs:
-      self.parse_mdn('sim_pose', outs, in_N=0, out_N=0, out_shape=(ModelConstants.POSE_WIDTH,))
+      self.parse_mdn('sim_pose', outs, out_shape=(ModelConstants.POSE_WIDTH,))
-    self.parse_mdn('wide_from_device_euler', outs, in_N=0, out_N=0, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
+    self.parse_mdn('wide_from_device_euler', outs, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
-    self.parse_mdn('lead', outs, in_N=ModelConstants.LEAD_MHP_N, out_N=ModelConstants.LEAD_MHP_SELECTION,
+    self.parse_mdn('lead', outs, out_shape=(ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH))
-                   out_shape=(ModelConstants.LEAD_TRAJ_LEN,ModelConstants.LEAD_WIDTH))
     if 'lat_planner_solution' in outs:
-      self.parse_mdn('lat_planner_solution', outs, in_N=0, out_N=0, out_shape=(ModelConstants.IDX_N,ModelConstants.LAT_PLANNER_SOLUTION_WIDTH))
+      self.parse_mdn('lat_planner_solution', outs, out_shape=(ModelConstants.IDX_N, ModelConstants.LAT_PLANNER_SOLUTION_WIDTH))
     if 'desired_curvature' in outs:
-      self.parse_mdn('desired_curvature', outs, in_N=0, out_N=0, out_shape=(ModelConstants.DESIRED_CURV_WIDTH,))
+      self.parse_mdn('desired_curvature', outs, out_shape=(ModelConstants.DESIRED_CURV_WIDTH,))
     for k in ['lead_prob', 'lane_lines_prob', 'meta']:
       self.parse_binary_crossentropy(k, outs)
     self.parse_categorical_crossentropy('desire_state', outs, out_shape=(ModelConstants.DESIRE_PRED_WIDTH,))
-    self.parse_categorical_crossentropy('desire_pred', outs, out_shape=(ModelConstants.DESIRE_PRED_LEN,ModelConstants.DESIRE_PRED_WIDTH))
+    self.parse_categorical_crossentropy('desire_pred',  outs, out_shape=(ModelConstants.DESIRE_PRED_LEN, ModelConstants.DESIRE_PRED_WIDTH))
     return outs

openpilot/sunnypilot/modeld_v2/parse_model_outputs_split.py

@@ -123,7 +123,7 @@ class Parser:
       self.parse_categorical_crossentropy('desire_state', outs, out_shape=(SplitModelConstants.DESIRE_PRED_WIDTH,))
     if 'lane_lines' in outs:
       self.parse_mdn('lane_lines', outs, in_N=0, out_N=0,
-                    out_shape=(SplitModelConstants.NUM_LANE_LINES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
+                     out_shape=(SplitModelConstants.NUM_LANE_LINES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
     if 'lane_lines_prob' in outs:
       self.parse_binary_crossentropy('lane_lines_prob', outs)
     if 'lead_prob' in outs:
@@ -134,7 +134,7 @@ class Parser:
       self.parse_binary_crossentropy('meta', outs)
     if 'road_edges' in outs:
       self.parse_mdn('road_edges', outs, in_N=0, out_N=0,
-                    out_shape=(SplitModelConstants.NUM_ROAD_EDGES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
+                     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:

openpilot/sunnypilot/modeld_v2/tests/test_combined_pkl_loader.py

@@ -67,16 +67,12 @@ class TestStockEquivalence:
     state = model_state_factory(ARCHETYPES['vision_policy_split'])
 
     frame_skip = derive_frame_skip(SPLIT_VISION_INPUT_SHAPES, SPLIT_POLICY_INPUT_SHAPES)
-    # action_t is a deep-model prerequisite the SP loader doesn't provide yet; see skip_keys below
     stock_shapes = {**SPLIT_VISION_INPUT_SHAPES, **SPLIT_POLICY_INPUT_SHAPES, 'action_t': (1, 2)}
     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)
+    assert set(state.input_queues.keys()) == set(stock_queues.keys())
-    skip_keys = {'action_t'}
+    assert {'desire', 'traffic_convention'} <= set(state.numpy_inputs.keys())
-    assert set(state.input_queues.keys()) == set(stock_queues.keys()) - skip_keys, \
+    assert set(state.numpy_inputs.keys()) == set(stock_npy.keys()) - {'action_t', 'prev_feat'}
-      f"Queue keys differ: v2={set(state.input_queues.keys())}, stock={set(stock_queues.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())}"
 
   def test_split_queue_keys_work_with_desire_key(self, model_state_factory):
     from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues

openpilot/sunnypilot/models/fetcher.py

@@ -13,6 +13,7 @@ from openpilot.common.params import Params
 from openpilot.common.swaglog import cloudlog
 from openpilot.common.hardware.hw import Paths
 from openpilot.sunnypilot.models.helpers import is_bundle_version_compatible
+from openpilot.selfdrive.modeld.helpers import usbgpu_present
 
 from openpilot.cereal import custom
 
@@ -64,8 +65,6 @@ class ModelParser:
 
     model.type = model_data.get("type")
     model.artifact = ModelParser._parse_artifact(model_data.get("artifact", {}))
-    if metadata := model_data.get("metadata"):
-      model.metadata = ModelParser._parse_artifact(metadata)
     return model
 
   @staticmethod
@@ -105,11 +104,11 @@ class ModelParser:
 class ModelCache:
   """Handles caching of model data to avoid frequent remote fetches"""
 
-  def __init__(self, params: Params, cache_timeout: int = int(3600 * 1e9)):
+  def __init__(self, params: Params, cache_timeout: int = int(3600 * 1e9), suffix: str = ""):
     self.params = params
     self.cache_timeout = cache_timeout
-    self._LAST_SYNC_KEY = "ModelManager_LastSyncTime"
+    self._LAST_SYNC_KEY = f"ModelManager_LastSyncTime{suffix}"
-    self._CACHE_KEY = "ModelManager_ModelsCache"
+    self._CACHE_KEY = f"ModelManager_ModelsCache{suffix}"
 
   def _is_expired(self) -> bool:
     """Checks if the cache has expired"""
@@ -141,24 +140,28 @@ 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"
 
   def __init__(self, params: Params):
     self.params = params
-    self.model_cache = ModelCache(params)
     self.model_parser = ModelParser()
+    if usbgpu_present():
+      self.model_cache = ModelCache(params, suffix="_USBGPU")
+      self.model_url = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_usbgpu_v18.json"
+    else:
+      self.model_cache = ModelCache(params)
+      self.model_url = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v18.json"
 
   def _fetch_and_cache_models(self) -> list[custom.ModelManagerSP.ModelBundle] | None:
     """Fetches fresh model data from remote and updates cache.
     Returns None on transport errors. Raises on 404 and other fatal HTTP errors.
     """
     try:
-      response = requests.get(self.MODEL_URL, timeout=10)
+      response = requests.get(self.model_url, timeout=10)
 
       # Explicitly handle 404 differently
       if response.status_code == 404:
-        cloudlog.error(f"Models URL returned 404 Not Found: {self.MODEL_URL}")
+        cloudlog.error(f"Models URL returned 404 Not Found: {self.model_url}")
-        raise HTTPError(f"404 Not Found: {self.MODEL_URL}", response=response)
+        raise HTTPError(f"404 Not Found: {self.model_url}", response=response)
 
       # Raise for any other 4xx/5xx
       response.raise_for_status()
@@ -211,5 +214,3 @@ if __name__ == "__main__":
       # Print metadata details
       if model.artifact.chunks:
         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/helpers.py

@@ -18,7 +18,7 @@ from openpilot.sunnypilot.models.constants import Meta, MetaSimPose, MetaTombRai
 from openpilot.common.hardware.hw import Paths
 
 # SET ME TO THE EXACT JSON VERSION WE SET IN SUNNYPILOT_MODELS REPO
-REQUIRED_JSON_VERSION = 15
+REQUIRED_JSON_VERSION = 16
 
 CUSTOM_MODEL_PATH = Paths.model_root()
 METADATA_PATH = Path(__file__).parent / '../models/supercombo_metadata.pkl'
@@ -56,12 +56,20 @@ def is_bundle_version_compatible(bundle: dict) -> bool:
 
 def _bundle_artifacts(bundle: custom.ModelManagerSP.ModelBundle) -> list[tuple[str, str]]:
   artifacts = []
+  from openpilot.common.file_chunker import get_chunk_name
   for model in getattr(bundle, 'models', []) or []:
-    for artifact in (getattr(model, 'artifact', None), getattr(model, 'metadata', None)):
+    for artifact in (getattr(model, 'artifact', None),):
-      if artifact and getattr(artifact, 'fileName', None) and getattr(artifact, 'downloadUri', None):
+      if artifact and getattr(artifact, 'fileName', None):
-        sha256 = getattr(artifact.downloadUri, 'sha256', None)
+        if len(artifact.chunks) > 0:
-        if sha256:
+          for i, chunk in enumerate(artifact.chunks):
-          artifacts.append((artifact.fileName, sha256))
+            chunk_name = get_chunk_name(artifact.fileName, i, len(artifact.chunks))
+            if getattr(chunk, 'sha256', None):
+              artifacts.append((chunk_name, chunk.sha256))
+        else:
+          if getattr(artifact, 'downloadUri', None):
+            sha256 = getattr(artifact.downloadUri, 'sha256', None)
+            if sha256:
+              artifacts.append((artifact.fileName, sha256))
   return artifacts
 
 
@@ -156,8 +164,7 @@ def _get_model():
 
 
 def load_metadata():
-  model = _get_model()
+  metadata_path = METADATA_PATH
-  metadata_path = f"{CUSTOM_MODEL_PATH}/{model.metadata.fileName}" if model else METADATA_PATH
 
   with open(metadata_path, 'rb') as f:
     return pickle.load(f)

openpilot/sunnypilot/models/manager.py

@@ -38,11 +38,11 @@ class ModelManagerSP:
     if not self.selected_bundle:
       return
     for model in self.selected_bundle.models:
-      for artifact in (model.artifact, model.metadata):
+      artifact = model.artifact
-        if artifact is not source_artifact and artifact.fileName == source_artifact.fileName:
+      if artifact is not source_artifact and artifact.fileName == source_artifact.fileName:
-          artifact.downloadProgress.status = source_artifact.downloadProgress.status
+        artifact.downloadProgress.status = source_artifact.downloadProgress.status
-          artifact.downloadProgress.progress = source_artifact.downloadProgress.progress
+        artifact.downloadProgress.progress = source_artifact.downloadProgress.progress
-          artifact.downloadProgress.eta = source_artifact.downloadProgress.eta
+        artifact.downloadProgress.eta = source_artifact.downloadProgress.eta
 
   def _calculate_eta(self, filename: str, progress: float) -> int:
     """Calculate ETA based on elapsed time and current progress"""
@@ -136,7 +136,22 @@ class ModelManagerSP:
     full_path = os.path.join(destination_path, filename)
 
     try:
-      if await verify_file(full_path, expected_hash):
+      is_cached = False
+      if len(artifact.chunks) > 0:
+        from openpilot.common.file_chunker import get_chunk_name
+        chunks_valid = True
+        for i, chunk in enumerate(artifact.chunks):
+          chunk_path = get_chunk_name(full_path, i, len(artifact.chunks))
+          if not await verify_file(chunk_path, chunk.sha256):
+            chunks_valid = False
+            break
+        if chunks_valid and len(artifact.chunks) > 0:
+          is_cached = True
+      else:
+        if await verify_file(full_path, expected_hash):
+          is_cached = True
+
+      if is_cached:
         artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.cached
         artifact.downloadProgress.progress = 100
         artifact.downloadProgress.eta = 0
@@ -146,11 +161,15 @@ class ModelManagerSP:
 
       if len(artifact.chunks) > 0:
         await self._download_chunked(url, full_path, artifact)
+        from openpilot.common.file_chunker import get_chunk_name
+        for i, chunk in enumerate(artifact.chunks):
+          chunk_path = get_chunk_name(full_path, i, len(artifact.chunks))
+          if not await verify_file(chunk_path, chunk.sha256):
+            raise ValueError(f"Hash validation failed for chunk {i+1} of {filename}")
       else:
         await self._download_file(url, full_path, artifact)
-
+        if not await verify_file(full_path, expected_hash):
-      if not await verify_file(full_path, expected_hash):
+          raise ValueError(f"Hash validation failed for {filename}")
-        raise ValueError(f"Hash validation failed for {filename}")
 
       artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloaded
       artifact.downloadProgress.progress = 100
@@ -198,16 +217,16 @@ class ModelManagerSP:
     try:
       seen_artifacts: set[str] = set()
       for model in self.selected_bundle.models:
-        for artifact in (model.metadata, model.artifact):
+        artifact = model.artifact
-          if not artifact.fileName:
+        if not artifact.fileName:
-            continue
+          continue
-          if artifact.fileName in seen_artifacts:
+        if artifact.fileName in seen_artifacts:
-            artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.cached
+          artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.cached
-            artifact.downloadProgress.progress = 100
+          artifact.downloadProgress.progress = 100
-            artifact.downloadProgress.eta = 0
+          artifact.downloadProgress.eta = 0
-          else:
+        else:
-            seen_artifacts.add(artifact.fileName)
+          seen_artifacts.add(artifact.fileName)
-            await self._process_artifact(artifact, destination_path)
+          await self._process_artifact(artifact, destination_path)
 
       self.active_bundle = self.selected_bundle
       self.active_bundle.status = custom.ModelManagerSP.DownloadStatus.downloaded
@@ -268,8 +287,6 @@ class ModelManagerSP:
       for model in self.active_bundle.models:
         if hasattr(model, 'artifact') and model.artifact.fileName:
           active_files.append(model.artifact.fileName)
-        if hasattr(model, 'metadata') and model.metadata.fileName:
-          active_files.append(model.metadata.fileName)
 
     # Remove all files except active ones (including their chunk files)
     model_dir = Paths.model_root()

openpilot/sunnypilot/models/tests/test_tinygrad_ref.py

@@ -1,11 +1,12 @@
 import requests
 
+from openpilot.common.params import Params
 from openpilot.sunnypilot.models.tinygrad_ref import get_tinygrad_ref
 from openpilot.sunnypilot.models.fetcher import ModelFetcher
 
-
 def fetch_tinygrad_ref():
-  response = requests.get(ModelFetcher.MODEL_URL, timeout=10)
+  fetcher = ModelFetcher(Params())
+  response = requests.get(fetcher.model_url, timeout=10)
   response.raise_for_status()
   json_data = response.json()
   return json_data.get("tinygrad_ref")

openpilot/sunnypilot/selfdrive/car/interfaces.py

@@ -123,6 +123,7 @@ def initialize_params(params) -> list[dict[str, Any]]:
   # tesla
   keys.extend([
     "TeslaCoopSteering",
+    "TeslaMadsScreenButton",
   ])
 
   # toyota

openpilot/sunnypilot/selfdrive/controls/lib/speed_limit/speed_limit_assist.py

@@ -91,7 +91,7 @@ class SpeedLimitAssist:
 
     self._plus_hold = 0.
     self._minus_hold = 0.
-    self._last_carstate_ts = 0.
+    self._release_toggle_prev = 0
 
     # TODO-SP: SLA's own output_a_target for planner
     # Solution functions mapped to respective states
@@ -146,16 +146,16 @@ class SpeedLimitAssist:
       set_speed_limit_assist_availability(self.CP, self.CP_SP, self.params)
       self.enabled = self.params.get("SpeedLimitMode", return_default=True) == Mode.assist
 
-  def update_car_state(self, CS: car.CarState) -> None:
+  def update_buttons(self, release_toggle: int) -> None:
+    released = self._release_toggle_prev ^ release_toggle
+    self._release_toggle_prev = release_toggle
+    if not released:
+      return
     now = time.monotonic()
-    self._last_carstate_ts = now
+    if any((released >> b) & 1 for b in CRUISE_BUTTONS_PLUS):
-
+      self._plus_hold = max(self._plus_hold, now + CRUISE_BUTTON_CONFIRM_HOLD)
-    for b in CS.buttonEvents:
+    if any((released >> b) & 1 for b in CRUISE_BUTTONS_MINUS):
-      if not b.pressed:
+      self._minus_hold = max(self._minus_hold, now + CRUISE_BUTTON_CONFIRM_HOLD)
-        if b.type in CRUISE_BUTTONS_PLUS:
-          self._plus_hold = max(self._plus_hold, now + CRUISE_BUTTON_CONFIRM_HOLD)
-        elif b.type in CRUISE_BUTTONS_MINUS:
-          self._minus_hold = max(self._minus_hold, now + CRUISE_BUTTON_CONFIRM_HOLD)
 
   def _get_button_release(self, req_plus: bool, req_minus: bool) -> bool:
     now = time.monotonic()

openpilot/sunnypilot/selfdrive/controls/lib/speed_limit/tests/test_speed_limit_assist.py

@@ -5,11 +5,14 @@ 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 time
+
 import pytest
 
 from openpilot.cereal import custom
 from opendbc.car.car_helpers import interfaces
 from opendbc.car.rivian.values import CAR as RIVIAN
+from opendbc.car.structs import car
 from opendbc.car.tesla.values import CAR as TESLA
 from opendbc.car.toyota.values import CAR as TOYOTA
 from openpilot.common.constants import CV
@@ -21,9 +24,13 @@ from openpilot.sunnypilot.selfdrive.car import interfaces as sunnypilot_interfac
 from openpilot.sunnypilot.selfdrive.controls.lib.speed_limit import PCM_LONG_REQUIRED_MAX_SET_SPEED
 from openpilot.sunnypilot.selfdrive.controls.lib.speed_limit.common import Mode
 from openpilot.sunnypilot.selfdrive.controls.lib.speed_limit.speed_limit_assist import SpeedLimitAssist, \
-  PRE_ACTIVE_GUARD_PERIOD, ACTIVE_STATES
+  PRE_ACTIVE_GUARD_PERIOD, ACTIVE_STATES, CRUISE_BUTTON_CONFIRM_HOLD
+from openpilot.sunnypilot.selfdrive.selfdrived.button_state_tracker import ButtonStateTracker
 from openpilot.sunnypilot.selfdrive.selfdrived.events import EventsSP
 
+ButtonEvent = car.CarState.ButtonEvent
+ButtonType = car.CarState.ButtonEvent.Type
+
 SpeedLimitAssistState = custom.LongitudinalPlanSP.SpeedLimit.AssistState
 
 ALL_STATES = tuple(SpeedLimitAssistState.schema.enumerants.values())
@@ -276,3 +283,86 @@ class TestSpeedLimitAssist:
         assert self.sla.state in [SpeedLimitAssistState.preActive, SpeedLimitAssistState.active]
       elif initial_state in ACTIVE_STATES:
         assert self.sla.state in ACTIVE_STATES
+
+
+class TestButtonStateTrackerSLAIntegration:
+
+  def setup_method(self, method):
+    self.tracker = ButtonStateTracker()
+    self.params = Params()
+    self.params.put("IsReleaseSpBranch", True, block=True)
+    self.params.put("SpeedLimitMode", int(Mode.assist), block=True)
+    self.params.put_bool("IsMetric", False, block=True)
+    self.params.put("SpeedLimitOffsetType", 0, block=True)
+    self.params.put("SpeedLimitValueOffset", 0, block=True)
+
+    CarInterface = interfaces[DEFAULT_CAR]
+    CP = CarInterface.get_non_essential_params(DEFAULT_CAR)
+    CP.openpilotLongitudinalControl = True
+    CP_SP = CarInterface.get_non_essential_params_sp(CP, DEFAULT_CAR)
+    self.sla = SpeedLimitAssist(CP, CP_SP)
+
+  def _make_cs(self, events=None) -> car.CarState:
+    CS = car.CarState()
+    CS.buttonEvents = events or []
+    return CS
+
+  def _run_ctrl_frames(self, frames: list[car.CarState]) -> None:
+    for cs in frames:
+      self.tracker.update(cs)
+
+  def test_button_confirm_via_tracker(self) -> None:
+    self._run_ctrl_frames([
+      self._make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=True)]),
+      self._make_cs(),
+      self._make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=False)]),
+      self._make_cs(),
+      self._make_cs(),
+    ])
+    self.sla.update_buttons(self.tracker.release_toggle)
+    assert self.sla._get_button_release(req_plus=True, req_minus=False)
+
+  def test_rapid_press_release_between_polls(self) -> None:
+    self.sla.update_buttons(self.tracker.release_toggle)
+
+    self._run_ctrl_frames([
+      self._make_cs([ButtonEvent(type=ButtonType.decelCruise, pressed=True)]),
+      self._make_cs([ButtonEvent(type=ButtonType.decelCruise, pressed=False)]),
+      self._make_cs(),
+      self._make_cs(),
+      self._make_cs(),
+    ])
+    self.sla.update_buttons(self.tracker.release_toggle)
+    assert self.sla._get_button_release(req_plus=False, req_minus=True)
+
+  def test_multiple_releases_between_polls(self) -> None:
+    self.sla.update_buttons(self.tracker.release_toggle)
+
+    self._run_ctrl_frames([
+      self._make_cs([
+        ButtonEvent(type=ButtonType.accelCruise, pressed=True),
+        ButtonEvent(type=ButtonType.decelCruise, pressed=True),
+      ]),
+      self._make_cs([
+        ButtonEvent(type=ButtonType.accelCruise, pressed=False),
+        ButtonEvent(type=ButtonType.decelCruise, pressed=False),
+      ]),
+    ])
+    self.sla.update_buttons(self.tracker.release_toggle)
+    assert self.sla._get_button_release(req_plus=True, req_minus=False)
+    assert self.sla._get_button_release(req_plus=False, req_minus=True)
+
+  def test_no_false_positive_same_toggle(self) -> None:
+    self.sla.update_buttons(self.tracker.release_toggle)
+    self.sla.update_buttons(self.tracker.release_toggle)
+    assert not self.sla._get_button_release(req_plus=True, req_minus=False)
+    assert not self.sla._get_button_release(req_plus=False, req_minus=True)
+
+  def test_button_confirm_expires(self) -> None:
+    self._run_ctrl_frames([
+      self._make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=True)]),
+      self._make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=False)]),
+    ])
+    self.sla.update_buttons(self.tracker.release_toggle)
+    time.sleep(CRUISE_BUTTON_CONFIRM_HOLD + 0.1)
+    assert not self.sla._get_button_release(req_plus=True, req_minus=False)

openpilot/sunnypilot/selfdrive/selfdrived/button_state_tracker.py

@@ -0,0 +1,26 @@
+"""
+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.
+"""
+from opendbc.car import structs
+
+
+class ButtonStateTracker:
+  def __init__(self) -> None:
+    self.pressed: int = 0
+    self.release_toggle: int = 0
+
+  def update(self, CS: structs.CarState) -> None:
+    for b in CS.buttonEvents:
+      bit = 1 << b.type.raw
+      if b.pressed:
+        self.pressed |= bit
+      else:
+        self.pressed &= ~bit
+        self.release_toggle ^= bit
+
+  def publish(self, ss_sp) -> None:
+    ss_sp.buttonsPressed = self.pressed
+    ss_sp.buttonsReleaseToggle = self.release_toggle

openpilot/sunnypilot/selfdrive/selfdrived/tests/test_button_state_tracker.py

@@ -0,0 +1,67 @@
+"""
+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.
+"""
+from opendbc.car.structs import car
+from openpilot.sunnypilot.selfdrive.selfdrived.button_state_tracker import ButtonStateTracker
+
+ButtonEvent = car.CarState.ButtonEvent
+ButtonType = car.CarState.ButtonEvent.Type
+
+
+class TestButtonStateTracker:
+  def setup_method(self) -> None:
+    self.tracker = ButtonStateTracker()
+
+  def make_cs(self, events: list) -> car.CarState:
+    CS = car.CarState()
+    CS.buttonEvents = events
+    return CS
+
+  def test_initial_state(self) -> None:
+    assert self.tracker.pressed == 0
+    assert self.tracker.release_toggle == 0
+
+  def test_press_sets_bit(self) -> None:
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=True)]))
+    assert self.tracker.pressed == (1 << ButtonType.accelCruise)
+    assert self.tracker.release_toggle == 0
+
+  def test_release_clears_and_toggles(self) -> None:
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=True)]))
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=False)]))
+    assert self.tracker.pressed == 0
+    assert self.tracker.release_toggle == (1 << ButtonType.accelCruise)
+
+  def test_multiple_buttons(self) -> None:
+    self.tracker.update(self.make_cs([
+      ButtonEvent(type=ButtonType.accelCruise, pressed=True),
+      ButtonEvent(type=ButtonType.decelCruise, pressed=True),
+    ]))
+    assert self.tracker.pressed == (1 << ButtonType.accelCruise) | (1 << ButtonType.decelCruise)
+
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=False)]))
+    assert self.tracker.pressed == (1 << ButtonType.decelCruise)
+    assert self.tracker.release_toggle == (1 << ButtonType.accelCruise)
+
+  def test_release_toggle_flips(self) -> None:
+    for _ in range(2):
+      self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.gapAdjustCruise, pressed=True)]))
+      self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.gapAdjustCruise, pressed=False)]))
+    assert self.tracker.release_toggle == 0
+
+  def test_publish(self) -> None:
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=True)]))
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.decelCruise, pressed=True)]))
+    self.tracker.update(self.make_cs([ButtonEvent(type=ButtonType.accelCruise, pressed=False)]))
+
+    class MockSP:
+      buttonsPressed = 0
+      buttonsReleaseToggle = 0
+
+    sp = MockSP()
+    self.tracker.publish(sp)
+    assert sp.buttonsPressed == self.tracker.pressed
+    assert sp.buttonsReleaseToggle == self.tracker.release_toggle

openpilot/sunnypilot/sunnylink/settings_ui.json

@@ -2161,6 +2161,42 @@
               "type": "offroad_only"
             }
           ]
+        },
+        {
+          "key": "TeslaMadsScreenButton",
+          "widget": "multiple_button",
+          "title": "MADS Screen Activation",
+          "description": "Use a multi-finger press on the infotainment screen to toggle MADS. This allows the use of full MADS functionality when enabled. Selecting a higher finger count may reduce accidental activations. Note: Setting this to Off will reset your MADS settings to default.",
+          "options": [
+            {
+              "value": 0,
+              "label": "Off"
+            },
+            {
+              "value": 1,
+              "label": "3-Finger"
+            },
+            {
+              "value": 2,
+              "label": "4-Finger"
+            },
+            {
+              "value": 3,
+              "label": "5-Finger"
+            }
+          ],
+          "visibility": [
+            {
+              "type": "capability",
+              "field": "tesla_has_vehicle_bus",
+              "equals": true
+            }
+          ],
+          "enablement": [
+            {
+              "type": "offroad_only"
+            }
+          ]
         }
       ]
     },

openpilot/sunnypilot/sunnylink/settings_ui_src/pages/vehicle.yaml

@@ -56,6 +56,28 @@ sections:
     title: Cooperative Steering (Beta)
     enablement:
     - $ref: '#/macros/offroad'
+  - key: TeslaMadsScreenButton
+    widget: multiple_button
+    title: MADS Screen Activation
+    description: 'Use a multi-finger press on the infotainment screen to toggle MADS.
+      This allows the use of full MADS functionality when enabled. Selecting a higher
+      finger count may reduce accidental activations. Note: Setting this to Off will
+      reset your MADS settings to default.'
+    options:
+    - value: 0
+      label: 'Off'
+    - value: 1
+      label: 3-Finger
+    - value: 2
+      label: 4-Finger
+    - value: 3
+      label: 5-Finger
+    visibility:
+    - type: capability
+      field: tesla_has_vehicle_bus
+      equals: true
+    enablement:
+    - $ref: '#/macros/offroad'
 - id: toyota
   title: Toyota / Lexus Settings
   description: ''

openpilot/sunnypilot/system/params_migration.py

@@ -17,6 +17,26 @@ ONROAD_BRIGHTNESS_TIMER_VALUES = {0: 3, 1: 5, 2: 7, 3: 10, 4: 15, 5: 30, **{i: (
 VALID_TIMER_VALUES = set(ONROAD_BRIGHTNESS_TIMER_VALUES.values())
 
 
+def _resolve_brand(_params) -> str:
+  bundle = _params.get("CarPlatformBundle")
+  if isinstance(bundle, dict) and bundle.get("brand"):
+    return str(bundle["brand"])
+
+  # Auto-fingerprinted cars have no bundle, fall back to the last known CarParams.
+  CP_bytes = _params.get("CarParamsPersistent")
+  if CP_bytes is None:
+    return ""
+
+  # Never raises: callers rely on "" to mean "brand unknown, skip the migration".
+  try:
+    from openpilot.cereal import messaging  # lazy: avoids heavy import at module level
+    from opendbc.car.structs import car
+    return str(messaging.log_from_bytes(CP_bytes, car.CarParams).brand)
+  except Exception as e:
+    cloudlog.exception(f"params_migration: failed to resolve brand from CarParamsPersistent: {e}")
+    return ""
+
+
 def _migrate_car_platform_bundle(_params):
   bundle = _params.get("CarPlatformBundle")
   if bundle is None:
@@ -47,6 +67,23 @@ def _migrate_car_platform_bundle(_params):
   cloudlog.info(f"params_migration: CarPlatformBundle migrated {old_platform!r} -> {new_platform!r}")
 
 
+def _migrate_tesla_mads_screen_button(_params):
+  # TeslaMadsScreenButton defaults to Off for fresh installs, but the screen button was previously always
+  # active on Teslas with a vehicle bus. Seed existing Tesla installs with 3-finger to preserve that.
+  try:
+    if _params.get("TeslaMadsScreenButton") is not None:
+      return
+
+    if _resolve_brand(_params) != "tesla":
+      return
+
+    from opendbc.sunnypilot.car.tesla.values import MadsScreenButtonType  # lazy: avoids heavy import at module level
+    _params.put("TeslaMadsScreenButton", MadsScreenButtonType.THREE_FINGER, block=True)
+    cloudlog.info("params_migration: seeded TeslaMadsScreenButton with 3-finger to preserve existing behavior")
+  except Exception as e:
+    cloudlog.exception(f"Error migrating TeslaMadsScreenButton: {e}")
+
+
 def run_migration(_params):
   # migrate OnroadScreenOffBrightness
   if _params.get("OnroadScreenOffBrightnessMigrated") != ONROAD_BRIGHTNESS_MIGRATION_VERSION:
@@ -80,3 +117,6 @@ def run_migration(_params):
       cloudlog.exception(f"Error migrating OnroadScreenOffTimer: {e}")
 
   _migrate_car_platform_bundle(_params)
+
+  # seed TeslaMadsScreenButton for existing Tesla installs
+  _migrate_tesla_mads_screen_button(_params)

release/ci/docker_build_sp.sh

@@ -1,30 +0,0 @@
-#!/usr/bin/env bash
-set -e
-
-SCRIPT_DIR=$(dirname "$0")
-OPENPILOT_DIR=$SCRIPT_DIR/../../
-
-DOCKER_IMAGE=sunnypilot
-DOCKER_FILE=Dockerfile.openpilot
-DOCKER_REGISTRY=ghcr.io/sunnypilot
-COMMIT_SHA=$(git rev-parse HEAD)
-
-if [ -n "$TARGET_ARCHITECTURE" ]; then
-  PLATFORM="linux/$TARGET_ARCHITECTURE"
-  TAG_SUFFIX="-$TARGET_ARCHITECTURE"
-else
-  PLATFORM="linux/$(uname -m)"
-  TAG_SUFFIX=""
-fi
-
-LOCAL_TAG=$DOCKER_IMAGE$TAG_SUFFIX
-REMOTE_TAG=$DOCKER_REGISTRY/$LOCAL_TAG
-REMOTE_SHA_TAG=$DOCKER_REGISTRY/$LOCAL_TAG:$COMMIT_SHA
-
-DOCKER_BUILDKIT=1 docker buildx build --provenance false --pull --platform $PLATFORM --load -t $DOCKER_IMAGE:latest -t $REMOTE_TAG -t $LOCAL_TAG -f $OPENPILOT_DIR/$DOCKER_FILE $OPENPILOT_DIR
-
-if [ -n "$PUSH_IMAGE" ]; then
-  docker push $REMOTE_TAG
-  docker tag $REMOTE_TAG $REMOTE_SHA_TAG
-  docker push $REMOTE_SHA_TAG
-fi