Force Stop: Fast Approach

Force Stop is now better at faster approach speeds. It has more authority on braking in general.

Force Stop Distance Offset: still the lever that moves the ending point sooner or later.

Radar bug fixed

cereal/custom.capnp

@@ -221,6 +221,11 @@ struct StarPilotPlan @0xf98d843bfd7004a3 {
   trackingLead @36 :Bool;
   stopSignConfirmed @37 :Bool;
   pulseGlideCoasting @38 :Bool;  # developer-only P&G phase for on-road status UI
+  # Curve Speed Controller diagnostics, for tuning and rollout validation
+  cscOverridden @39 :Bool;          # driver cancelled this curve with RES+
+  cscLearnedLatAccel @40 :Float32;  # learned comfort at the current curvature, before margin
+  cscBindingDistance @41 :Float32;  # distance to the horizon point setting the target, m
+  approachStopLength @42 :Float32;  # pre-commit distance to a detected stop, m; 0 when off
 }
 
 struct StarPilotRadarState @0xb86e6369214c01c8 {

selfdrive/controls/controlsd.py

@@ -169,12 +169,27 @@ CURVATURE_HOLD_OPPOSITE_RELEASE = 0.01  # 1/m
 CURVATURE_HOLD_CONFIRM_MIN = 0.003  # 1/m (~7 deg) of wound curvature before capture
 CURVATURE_HOLD_CONFIRM_SWEPT = 0.6  # rad of heading swept this blinker cycle; past this the push is exit-shaping, not initiation
 
+# Suppress low-speed action spikes while the model's spatial path remains straight.
+TWITCH_GUARD_MAX_SPEED = 4.0
+TWITCH_GUARD_FADE_SPEED = 3.0
+TWITCH_GUARD_DURATION = 1.5
+TWITCH_GUARD_PLAN_RATIO = 4.0
+TWITCH_GUARD_FLOOR = 0.002
+TWITCH_GUARD_STRAIGHT_LO = 0.005
+TWITCH_GUARD_STRAIGHT_HI = 0.014
+TWITCH_GUARD_MIN_REACH = 12.0
+
 
 def _plan_circle_curvature(xs, ys, lookahead: float) -> float:
-  # curvature of the circle through the origin, tangent to the car's heading, passing
-  # through the plan point ~lookahead meters ahead: kappa = 2y / (x^2 + y^2)
+  # Fit curvature through the plan point at the requested lookahead.
   px, py = 0.0, 0.0
-  for x, y in zip(xs, ys):
+  for x, y in zip(xs, ys, strict=False):
+    try:
+      x, y = float(x), float(y)
+    except (TypeError, ValueError, OverflowError):
+      return 0.0
+    if not (math.isfinite(x) and math.isfinite(y)):
+      return 0.0
     px, py = x, y
     if math.hypot(x, y) >= lookahead:
       break
@@ -185,11 +200,7 @@ def _plan_circle_curvature(xs, ys, lookahead: float) -> float:
 
 
 def _plan_dual_probe(model_v2, d_near: float, d_far: float) -> float:
-  # Min-magnitude of a near and a far circle fit. The far probe alone assumes the turn
-  # starts immediately, which over-winds wide turns whose arc begins several meters out
-  # (wide multi-lane lefts): the near probe reads ~straight there and only grows as the
-  # car approaches the arc, so the readout self-scales to the turn geometry. Sign
-  # disagreement means no coherent turn ahead: contribute nothing.
+  # Use the smaller magnitude of near and far probes to avoid early turn bias.
   xs, ys = model_v2.position.x, model_v2.position.y
   near = _plan_circle_curvature(xs, ys, d_near)
   far = _plan_circle_curvature(xs, ys, d_far)
@@ -222,8 +233,52 @@ def get_plan_turn_onset_dist(model_v2) -> float:
 
 
 def get_plan_reach(model_v2) -> float:
-  xs = model_v2.position.x
-  return xs[-1] if len(xs) else 0.0
+  try:
+    xs = model_v2.position.x
+    return float(xs[-1]) if len(xs) else 0.0
+  except (AttributeError, IndexError, TypeError, ValueError, OverflowError):
+    return 0.0
+
+
+def _plan_positions_are_finite(model_v2) -> bool:
+  try:
+    xs, ys = model_v2.position.x, model_v2.position.y
+    return len(xs) == len(ys) and all(
+      math.isfinite(float(x)) and math.isfinite(float(y)) for x, y in zip(xs, ys, strict=True)
+    )
+  except (AttributeError, TypeError, ValueError, OverflowError):
+    return False
+
+
+def limit_curvature_to_plan(model_v2, curvature: float, v_ego: float) -> float:
+  if not (math.isfinite(curvature) and math.isfinite(v_ego)):
+    return curvature
+  if v_ego >= TWITCH_GUARD_MAX_SPEED or curvature == 0.0:
+    return curvature
+  if not _plan_positions_are_finite(model_v2):
+    return curvature
+  reach = get_plan_reach(model_v2)
+  if not math.isfinite(reach) or reach < TWITCH_GUARD_MIN_REACH:
+    return curvature
+  plan = abs(_plan_circle_curvature(model_v2.position.x, model_v2.position.y,
+                                    CURVATURE_HOLD_PLAN_LOOKAHEAD_FAR))
+  if not math.isfinite(plan):
+    return curvature
+  straightness = (plan - TWITCH_GUARD_STRAIGHT_LO) / (TWITCH_GUARD_STRAIGHT_HI - TWITCH_GUARD_STRAIGHT_LO)
+  limit = max(TWITCH_GUARD_PLAN_RATIO * plan * min(max(straightness, 0.0), 1.0), TWITCH_GUARD_FLOOR)
+  if abs(curvature) <= limit:
+    return curvature
+  fade = (TWITCH_GUARD_MAX_SPEED - v_ego) / (TWITCH_GUARD_MAX_SPEED - TWITCH_GUARD_FADE_SPEED)
+  fade = min(max(fade, 0.0), 1.0)
+  return curvature + (math.copysign(limit, curvature) - curvature) * fade
+
+
+def update_twitch_guard(remaining: float, v_ego: float, standstill: bool) -> float:
+  if not (math.isfinite(remaining) and math.isfinite(v_ego)):
+    return 0.0
+  if standstill or abs(v_ego) <= 0.3:
+    return TWITCH_GUARD_DURATION
+  return max(remaining - DT_CTRL, 0.0)
 
 
 def get_control_lateral_smooth_seconds(brand: str, v_ego: float, vehicle_smooth_seconds: float) -> float:
@@ -346,6 +401,7 @@ class Controls:
     self.turn_hold_handoff_t = 0.0
     self.turn_hold_done = False
     self.turn_blinker_swept = 0.0
+    self.twitch_guard_remaining = 0.0
     self.kona_non_scc_lateral_active = False
 
     self.pose_calibrator = PoseCalibrator()
@@ -401,6 +457,7 @@ class Controls:
 
   def state_control(self):
     CS = self.sm['carState']
+    self.twitch_guard_remaining = update_twitch_guard(self.twitch_guard_remaining, CS.vEgo, CS.standstill)
 
     # Update VehicleModel
     lp = self.sm['liveParameters']
@@ -460,7 +517,11 @@ class Controls:
     # EcuDisableFailed is set when car started in READY mode (ECU disable was rejected)
     # Disable longitudinal so stock ACC works instead
     self.update_ecu_disable_failed()
-    CC.longActive = CC.enabled and not any(e.overrideLongitudinal for e in self.sm['onroadEvents']) and not self.sm['starpilotCarState'].pauseLongitudinal and self.CP.openpilotLongitudinalControl and not self.ecu_disable_failed
+    CC.longActive = (
+      CC.enabled and not any(e.overrideLongitudinal for e in self.sm['onroadEvents']) and
+      not self.sm['starpilotCarState'].pauseLongitudinal and self.CP.openpilotLongitudinalControl and
+      not self.ecu_disable_failed
+    )
 
     actuators = CC.actuators
     actuators.longControlState = self.LoC.long_control_state
@@ -499,6 +560,9 @@ class Controls:
     # here is positive for RIGHT turns (pauseturn log: left turn at +148 deg steering
     # angle logs desiredCurvature -0.07), so the blinker maps right=+1, left=-1.
     blinker_dir = float(CS.rightBlinker) - float(CS.leftBlinker)
+    if (CC.latActive and self.twitch_guard_remaining > 0.0 and
+        blinker_dir == 0.0 and self.turn_hold_curvature == 0.0):
+      new_desired_curvature = limit_curvature_to_plan(model_v2, new_desired_curvature, CS.vEgo)
     # heading swept in the blinker's direction over the whole blinker cycle (any speed):
     # discriminates a turn not yet made from one being exited (see the re-arm below)
     if blinker_dir == 0.0:

selfdrive/controls/lib/lane_centering.py

@@ -85,7 +85,8 @@ class LaneCenteringController:
   def _covers(x, distance: float) -> bool:
     return bool(x[0] <= distance <= x[-1])
 
-  def _raw_correction(self, model_v2, v_ego: float, offset: float, e2e_authority: float) -> tuple[bool, float]:
+  @staticmethod
+  def _raw_correction(model_v2, v_ego: float, offset: float, e2e_authority: float) -> tuple[bool, float]:
     try:
       lane_lines = model_v2.laneLines
       probs = np.asarray(model_v2.laneLineProbs, dtype=float)
@@ -105,11 +106,13 @@ class LaneCenteringController:
       right_y = np.asarray(lane_lines[2].y, dtype=float)
       pos_x = np.asarray(model_v2.position.x, dtype=float)
       pos_y = np.asarray(model_v2.position.y, dtype=float)
-      if not (self._valid_path(left_x, left_y) and self._valid_path(right_x, right_y) and self._valid_path(pos_x, pos_y)):
+      if not (LaneCenteringController._valid_path(left_x, left_y) and
+              LaneCenteringController._valid_path(right_x, right_y) and
+              LaneCenteringController._valid_path(pos_x, pos_y)):
         return False, 0.0
 
       lookahead = float(np.clip(v_ego, 8.0, 35.0))
-      if not all(self._covers(x, lookahead) for x in (left_x, right_x, pos_x)):
+      if not all(LaneCenteringController._covers(x, lookahead) for x in (left_x, right_x, pos_x)):
         return False, 0.0
 
       left = float(np.interp(lookahead, left_x, left_y))
@@ -130,7 +133,7 @@ class LaneCenteringController:
 
       try:
         pos_y_std = np.asarray(model_v2.position.yStd, dtype=float)
-        if self._valid_path(pos_x, pos_y_std):
+        if LaneCenteringController._valid_path(pos_x, pos_y_std):
           path_std = float(np.interp(lookahead, pos_x, pos_y_std))
           if 0.0 <= path_std <= _E2E_MAX_PATH_STD:
             break_in = np.clip(
@@ -145,3 +148,43 @@ class LaneCenteringController:
       return True, float(2.0 * error / lookahead ** 2)
     except (AttributeError, IndexError, TypeError, ValueError):
       return False, 0.0
+
+
+def get_raw_lane_centering_correction(model_v2, v_ego: float, offset: float,
+                                      e2e_authority: float) -> tuple[bool, float]:
+  """Return the instantaneous lane-centering correction without controller filtering."""
+  return LaneCenteringController._raw_correction(model_v2, v_ego, offset, e2e_authority)
+
+
+def get_lane_centering_visual_direction(model_v2, v_ego: float, offset: float, e2e_authority: float,
+                                        enabled: bool, lat_active: bool, pause_on_signal: bool = False,
+                                        turn_signal_active: bool = False,
+                                        applied_correction: float | None = None) -> int:
+  """Return 1 for a right correction, -1 for left, and 0 when no correction is active."""
+  if not enabled or not lat_active or (pause_on_signal and turn_signal_active):
+    return 0
+
+  try:
+    v_ego = float(v_ego)
+    offset = float(offset)
+    e2e_authority = float(e2e_authority)
+    if not np.isfinite([v_ego, offset, e2e_authority]).all() or v_ego < _MIN_V_EGO:
+      return 0
+    if model_v2.meta.laneChangeState != log.LaneChangeState.off:
+      return 0
+  except (AttributeError, TypeError, ValueError):
+    return 0
+
+  valid, correction = get_raw_lane_centering_correction(
+    model_v2,
+    v_ego,
+    float(np.clip(offset, -_MAX_OFFSET, _MAX_OFFSET)),
+    float(np.clip(e2e_authority, 0.0, 1.0)),
+  )
+  if not valid or not np.isfinite(correction):
+    return 0
+  if correction == 0.0:
+    if applied_correction is None or not np.isfinite(applied_correction) or applied_correction == 0.0:
+      return 0
+    correction = applied_correction
+  return 1 if correction > 0.0 else -1

selfdrive/controls/lib/longitudinal_planner.py

@@ -9,6 +9,7 @@ from openpilot.common.filter_simple import FirstOrderFilter
 from openpilot.common.realtime import DT_MDL
 from openpilot.selfdrive.modeld.constants import ModelConstants
 from openpilot.starpilot.common.model_versions import is_tinygrad_model_version
+from openpilot.starpilot.controls.lib.starpilot_vcruise import FT_TO_M, OFFSET_FT_MAX, OFFSET_FT_MIN
 from openpilot.selfdrive.controls.lib.longcontrol import LongCtrlState
 from openpilot.selfdrive.controls.lib.longitudinal_mpc_lib.long_mpc import LongitudinalMpc
 from openpilot.selfdrive.controls.lib.longitudinal_mpc_lib.long_mpc import desired_follow_distance
@@ -381,6 +382,11 @@ def get_vehicle_min_accel(CP, v_ego):
 
 # Restored planner constants retained by CEM, stop, and departure paths.
 A_CRUISE_MIN = -1.0
+# The stop distance runs ~9 m long through the mid-approach, which leaves the obstacle slack
+# so it stays silent and deceleration sags. Multiplicative so the trim scales with what is
+# left. Note the car parks where the obstacle sits, so this is also a placement bias — 0.85
+# stopped ~4.6 m short, 0.93 ~1.6 m.
+FORCE_STOP_OBSTACLE_TRIM = 0.93
 STANDSTILL_LEAD_CREEP_RELEASE_MIN_LEAD_SPEED = 0.25
 STANDSTILL_LEAD_CREEP_RELEASE_MIN_LEAD_ACCEL = 0.08
 STANDSTILL_LEAD_CREEP_RELEASE_MIN_GAP_MARGIN = 0.1
@@ -2217,8 +2223,17 @@ class LongitudinalPlanner:
     force_stop_x = None
     force_stop_handoff_m = get_force_stop_handoff_distance(self.CP.carFingerprint)
     if sm['starpilotPlan'].forcingStop and sm['starpilotPlan'].forcingStopLength > force_stop_handoff_m:
+      stop_length = float(sm['starpilotPlan'].forcingStopLength)
+    else:
+      # pre-commit the envelope is only a speed ceiling, which the solver tracks with a lag;
+      # getattr so a stale cereal build degrades to the old behaviour instead of raising
+      stop_length = float(getattr(sm['starpilotPlan'], 'approachStopLength', 0.0))
+    if stop_length > force_stop_handoff_m:
+      # ForceStopDistanceOffset shifts the perceived line for the v_cruise ceiling, so it has
+      # to shift the obstacle too or the slider barely moves anything now that stop_x leads.
+      offset_ft = max(OFFSET_FT_MIN, min(OFFSET_FT_MAX, int(getattr(starpilot_toggles, 'force_stop_distance_offset', 0) or 0)))
       force_stop_x = (
-        float(sm['starpilotPlan'].forcingStopLength) + STOP_DISTANCE +
+        stop_length * FORCE_STOP_OBSTACLE_TRIM + offset_ft * FT_TO_M + STOP_DISTANCE +
         get_force_stop_distance_bias(self.CP.carFingerprint)
       )
 

selfdrive/controls/radard.py

@@ -53,6 +53,8 @@ ADJACENT_STOP_REST_FRAMES = 15
 ADJACENT_STOP_MIN_Y = 1.8         # m — inside this is our own lane
 ADJACENT_STOP_MAX_Y = 7.5         # m — beyond this is roadside, not an adjacent lane
 ADJACENT_STOP_MAX_D = 110.0       # m
+ADJACENT_STOP_QUEUE_GAP_M = 5.0   # m — anything stopped beyond the furthest qualifier means
+                                  # the bar is past it too, so the hint would stop us short
 
 
 class KalmanParams:
@@ -179,6 +181,10 @@ class Track:
     if self.leadTrackID == self.identifier:
       return False
 
+    return self.in_adjacent_lane(model_data)
+
+  def in_adjacent_lane(self, model_data: capnp._DynamicStructReader):
+    """Lane geometry only, no deceleration history — also used to spot a queue ahead."""
     if not (ADJACENT_STOP_MIN_Y < abs(self.yRel) < ADJACENT_STOP_MAX_Y):
       return False
 
@@ -398,9 +404,10 @@ def get_adjacent_lead(tracks: dict[int, Track], standstill: bool, model_data: ca
 def get_adjacent_stopped(tracks: dict[int, Track], model_data: capnp._DynamicStructReader) -> dict[str, Any]:
   """Stop-line hint: a vehicle that decelerated to a stop in a neighbouring lane.
 
-  Takes the FARTHEST qualifying vehicle: in a queue the front car sits at the bar and the
-  rest are closer to us, so the nearest one underestimates the distance. The consumer only
-  shortens with this, so underestimating is the harmful direction.
+  Takes the FARTHEST qualifying vehicle, then drops the hint entirely if a queue reaches
+  past it. Cars already stopped when we acquire them never show the moving -> stopped
+  transition, so the qualifying set is biased toward the back of a line; without this the
+  hint marks a mid-queue bumper and stops us short of the bar.
   """
   if len(model_data.laneLines) < 4:
     return {'status': False}
@@ -410,6 +417,11 @@ def get_adjacent_stopped(tracks: dict[int, Track], model_data: capnp._DynamicStr
     return {'status': False}
 
   furthest = max(candidates, key=lambda c: c.dRel)
+  for c in tracks.values():
+    if (c.dRel > furthest.dRel + ADJACENT_STOP_QUEUE_GAP_M and
+        abs(c.vLead) < ADJACENT_STOP_REST_V and
+        c.in_adjacent_lane(model_data)):
+      return {'status': False}
   return {
     'status': True,
     'dRel': float(furthest.dRel),

selfdrive/controls/tests/test_lane_centering.py

@@ -3,7 +3,7 @@ from types import SimpleNamespace
 import numpy as np
 import pytest
 
-from openpilot.selfdrive.controls.lib.lane_centering import LaneCenteringController
+from openpilot.selfdrive.controls.lib.lane_centering import LaneCenteringController, get_lane_centering_visual_direction
 
 
 _V_EGO = 20.0
@@ -174,3 +174,20 @@ def test_correction_is_smoothed_and_capped():
   _, steady = _converge(model, authority=0.0)
   assert 0.0 < first < steady
   assert np.isclose(steady, 0.004 * 0.30, atol=1e-6)
+
+
+def test_visual_direction_matches_curvature_sign():
+  # Positive curvature is right in this tree's convention.
+  assert get_lane_centering_visual_direction(_model(left=-1.5, right=2.1), _V_EGO, 0.0, 0.0, True, True) == 1
+  assert get_lane_centering_visual_direction(_model(left=-2.1, right=1.5), _V_EGO, 0.0, 0.0, True, True) == -1
+
+
+def test_visual_direction_requires_both_primary_lane_lines():
+  model = _model(left=-1.5, right=2.1)
+  model.laneLineProbs[2] = 0.2
+  assert get_lane_centering_visual_direction(model, _V_EGO, 0.0, 0.0, True, True) == 0
+
+
+def test_visual_direction_uses_filtered_correction_in_deadband():
+  model = _model()
+  assert get_lane_centering_visual_direction(model, _V_EGO, 0.0, 0.0, True, True, applied_correction=0.001) == 1

selfdrive/controls/tests/test_longitudinal_planner.py

@@ -526,6 +526,7 @@ def make_sm(v_ego: float, desired_accel: float, min_accel: float, *, experimenta
       forcingStop=False,
       redLight=False,
       forcingStopLength=2,
+      approachStopLength=0.0,
     ),
   }
 
@@ -2990,6 +2991,39 @@ def test_modeld_action_uses_current_action_head_scaling_for_v15(monkeypatch):
   assert not action.shouldStop
 
 
+def test_modeld_action_uses_current_action_head_scaling_for_v16(monkeypatch):
+  monkeypatch.setenv("DEBUG", "0")
+  fake_commonmodel = types.ModuleType("openpilot.selfdrive.modeld.models.commonmodel_pyx")
+  fake_commonmodel.DrivingModelFrame = object
+  fake_commonmodel.CLContext = object
+  monkeypatch.setitem(sys.modules, fake_commonmodel.__name__, fake_commonmodel)
+
+  from openpilot.selfdrive.modeld import modeld
+
+  prev_action = log.ModelDataV2.Action.new_message()
+  prev_action.desiredCurvature = 0.05
+  prev_action.desiredAcceleration = -0.2
+  toggles = SimpleNamespace(vEgoStopping=0.42)
+
+  action = modeld.get_action_from_model(
+    {"action": np.array([[12.0, -0.8]], dtype=np.float32)},
+    prev_action,
+    lat_action_t=0.2,
+    long_action_t=0.73,
+    v_ego=5.0,
+    mlsim=True,
+    is_v9=False,
+    is_v14=False,
+    is_v15=False,
+    starpilot_toggles=toggles,
+    is_v16=True,
+  )
+
+  assert action.desiredCurvature == pytest.approx(modeld.smooth_value(0.48, prev_action.desiredCurvature, modeld.LAT_SMOOTH_SECONDS))
+  assert action.desiredAcceleration < -0.2
+  assert not action.shouldStop
+
+
 def test_publish_force_stop_handoff_sets_should_stop_when_vcruise_zero():
   class FakePM:
     def __init__(self):

selfdrive/controls/tests/test_starpilot_vcruise.py

@@ -7,7 +7,9 @@ from openpilot.common.realtime import DT_MDL
 from openpilot.starpilot.common.starpilot_variables import PLANNER_TIME
 from openpilot.starpilot.controls.lib.curve_speed_controller import CSC_MAX_DECEL_RATE, CurveSpeedController
 from openpilot.starpilot.controls.lib.starpilot_vcruise import (
+  FORCE_STOP_CAP_SLACK_M,
   FORCE_STOP_TURN_VETO_STOP_SEEN_HOLD_TIME,
+  STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME,
   StarPilotVCruise,
   get_active_slc_control_target,
   get_lead_veto_distance,
@@ -57,6 +59,8 @@ def make_vcruise(*, red_light=False, raw_model_stopped=False, forcing_stop=False
   vcruise.forcing_stop = forcing_stop
   vcruise.force_stop_timer = 1.0 if forcing_stop else 0.0
   vcruise.tracked_model_length = 0.0 if forcing_stop else planner.model_length
+  # what the not-committed branch would have left behind on the frame before commit
+  vcruise.force_stop_distance_cap = planner.model_length
   return planner, vcruise
 
 
@@ -521,14 +525,15 @@ def test_force_stop_stays_committed_while_moving_even_if_scene_opens():
 
 def test_force_stop_reanchors_when_model_reopens_path_without_stop_action():
   planner, vcruise = make_vcruise(red_light=False, raw_model_stopped=False, forcing_stop=True)
-  planner.model_length = 40.0
-  vcruise.tracked_model_length = 10.0
+  planner.model_length = 90.0
+  vcruise.tracked_model_length = 60.0
+  vcruise.force_stop_distance_cap = 90.0
   sm = make_sm(standstill=False)
   sm["modelV2"] = SimpleNamespace(action=SimpleNamespace(shouldStop=False))
 
   result = update_vcruise(vcruise, sm, make_toggles(), now=0.0, v_ego=1.5)
 
-  assert vcruise.tracked_model_length == pytest.approx(40.0)
+  assert vcruise.tracked_model_length == pytest.approx(90.0)
   assert result > 5.0
 
 
@@ -560,6 +565,49 @@ def test_santa_fe_force_stop_holds_through_low_speed_detector_dropout():
   assert result == pytest.approx(0.0)
 
 
+def test_force_stop_does_not_reanchor_inside_reanchor_floor():
+  planner, vcruise = make_vcruise(red_light=False, raw_model_stopped=False, forcing_stop=True)
+  planner.model_length = 90.0
+  vcruise.tracked_model_length = 25.0
+  sm = make_sm(standstill=False)
+  sm["modelV2"] = SimpleNamespace(action=SimpleNamespace(shouldStop=False))
+
+  update_vcruise(vcruise, sm, make_toggles(), now=0.0, v_ego=1.5)
+
+  assert vcruise.tracked_model_length < 25.0
+
+
+def test_force_stop_reanchor_bounded_by_distance_driven():
+  # The line can't recede: a ballooning horizon may not push the stop past where it was at
+  # commit minus the distance driven since.
+  planner, vcruise = make_vcruise(red_light=False, raw_model_stopped=False, forcing_stop=True)
+  planner.model_length = 200.0
+  vcruise.tracked_model_length = 60.0
+  vcruise.force_stop_distance_cap = 70.0
+  sm = make_sm(standstill=False)
+  sm["modelV2"] = SimpleNamespace(action=SimpleNamespace(shouldStop=False))
+
+  update_vcruise(vcruise, sm, make_toggles(), now=0.0, v_ego=10.0)
+
+  assert vcruise.tracked_model_length <= 70.0 + FORCE_STOP_CAP_SLACK_M
+  assert vcruise.tracked_model_length < 100.0  # nowhere near the 200 m the horizon claimed
+
+
+def test_force_stop_cap_slack_tapers_near_the_line():
+  # Slack protects against an under-read at commit; held near the line it would just aim the
+  # solver that far past the stop bar.
+  planner, vcruise = make_vcruise(red_light=False, raw_model_stopped=False, forcing_stop=True)
+  planner.model_length = 200.0
+  vcruise.tracked_model_length = 60.0
+  vcruise.force_stop_distance_cap = 12.0
+  sm = make_sm(standstill=False)
+  sm["modelV2"] = SimpleNamespace(action=SimpleNamespace(shouldStop=False))
+
+  update_vcruise(vcruise, sm, make_toggles(), now=0.0, v_ego=5.0)
+
+  assert vcruise.tracked_model_length < 12.0 + FORCE_STOP_CAP_SLACK_M / 2.0
+
+
 def test_force_stop_does_not_reanchor_committed_model_stop():
   planner, vcruise = make_vcruise(red_light=False, raw_model_stopped=False, forcing_stop=True)
   planner.model_length = 40.0
@@ -841,16 +889,16 @@ def test_standstill_light_hold_expires_and_does_not_rearm_from_stopped_model():
   assert update_vcruise(vcruise, sm, toggles, now=0.0) == pytest.approx(0.0)
   assert vcruise.standstill_force_stop_reason == "light"
 
-  assert update_vcruise(vcruise, sm, toggles, now=4.9) == pytest.approx(0.0)
+  assert update_vcruise(vcruise, sm, toggles, now=STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME - 0.1) == pytest.approx(0.0)
   assert vcruise.forcing_stop
 
-  assert update_vcruise(vcruise, sm, toggles, now=5.1) == pytest.approx(20.0)
+  assert update_vcruise(vcruise, sm, toggles, now=STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME + 0.1) == pytest.approx(20.0)
   assert not vcruise.forcing_stop
   assert not vcruise.standstill_force_stop_hold
 
   # The red-light model remains stopped, but Force Stop must stay released so
   # Experimental Mode can own the red-to-green departure.
-  assert update_vcruise(vcruise, sm, toggles, now=5.2) == pytest.approx(20.0)
+  assert update_vcruise(vcruise, sm, toggles, now=STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME + 0.2) == pytest.approx(20.0)
   assert not vcruise.forcing_stop
 
 

selfdrive/controls/tests/test_turn_lead.py

@@ -1,13 +1,40 @@
+import math
+import types
+
 from cereal import car
 
 import pytest
 
-from openpilot.selfdrive.controls.controlsd import get_control_lateral_smooth_seconds, turn_lead_allowed
+from openpilot.common.realtime import DT_CTRL
+from openpilot.selfdrive.controls.controlsd import (
+  TWITCH_GUARD_DURATION,
+  TWITCH_GUARD_FLOOR,
+  TWITCH_GUARD_MAX_SPEED,
+  get_control_lateral_smooth_seconds,
+  limit_curvature_to_plan,
+  turn_lead_allowed,
+  update_twitch_guard,
+)
 
 
 LateralControlMode = car.CarControl.Actuators.LateralControlMode
 
 
+def _plan(xs, ys):
+  return types.SimpleNamespace(position=types.SimpleNamespace(x=xs, y=ys))
+
+
+def _arc_plan(radius, n=200):
+  return _plan([radius * math.sin(i / n) for i in range(n)],
+               [radius * (1.0 - math.cos(i / n)) for i in range(n)])
+
+
+STRAIGHT_PLAN = _plan([i * 0.5 for i in range(200)], [0.0] * 200)
+STANDSTILL_STUB_PLAN = _plan([0.0, 0.3], [0.0, 0.0])
+TURN_PLAN = _arc_plan(30.0)
+GENTLE_BEND_PLAN = _arc_plan(143.0)
+
+
 def test_turn_lead_is_suppressed_only_during_applied_angle_control():
   assert not turn_lead_allowed("rivian", LateralControlMode.angle)
   assert turn_lead_allowed("rivian", LateralControlMode.torque)
@@ -37,3 +64,79 @@ def test_subaru_control_smoothing_uses_vehicle_schedule(v_ego, expected):
 ])
 def test_rivian_control_smoothing_remains_speed_scheduled(v_ego, expected):
   assert get_control_lateral_smooth_seconds("rivian", v_ego, 0.4) == pytest.approx(expected)
+
+
+@pytest.mark.parametrize("curvature", [0.0155, -0.0155])
+def test_twitch_against_a_straight_plan_is_clamped_to_the_floor(curvature):
+  guarded = limit_curvature_to_plan(STRAIGHT_PLAN, curvature, 1.2)
+  assert abs(guarded) == pytest.approx(TWITCH_GUARD_FLOOR)
+  assert math.copysign(1.0, guarded) == math.copysign(1.0, curvature)
+
+
+def test_command_already_below_the_floor_is_untouched():
+  assert limit_curvature_to_plan(STRAIGHT_PLAN, 0.0015, 1.2) == pytest.approx(0.0015)
+
+
+@pytest.mark.parametrize("v_ego", [TWITCH_GUARD_MAX_SPEED, 6.0, 30.0])
+def test_guard_is_inactive_above_its_speed_band(v_ego):
+  assert limit_curvature_to_plan(STRAIGHT_PLAN, 0.0155, v_ego) == pytest.approx(0.0155)
+
+
+def test_guard_fades_out_across_the_speed_band():
+  full = limit_curvature_to_plan(STRAIGHT_PLAN, 0.0155, 1.2)
+  half = limit_curvature_to_plan(STRAIGHT_PLAN, 0.0155, 3.5)
+  assert full < half < 0.0155
+
+
+@pytest.mark.parametrize("ratio", [0.8, 1.0, 2.0, 3.0])
+def test_real_turns_tracking_their_own_plan_are_untouched(ratio):
+  action = (1.0 / 30.0) * ratio
+  assert limit_curvature_to_plan(TURN_PLAN, action, 1.2) == pytest.approx(action)
+
+
+def test_a_barely_bending_plan_does_not_license_a_large_command():
+  guarded = limit_curvature_to_plan(GENTLE_BEND_PLAN, 0.0155, 1.2)
+  assert TWITCH_GUARD_FLOOR < guarded < 0.008
+
+
+@pytest.mark.parametrize("plan", [STANDSTILL_STUB_PLAN, _plan([], [])])
+def test_guard_stands_down_when_the_plan_is_too_short_to_judge(plan):
+  assert limit_curvature_to_plan(plan, 0.0155, 0.4) == pytest.approx(0.0155)
+
+
+def test_zero_command_stays_zero():
+  assert limit_curvature_to_plan(STRAIGHT_PLAN, 0.0, 1.2) == 0.0
+
+
+@pytest.mark.parametrize("bad_plan", [
+  _plan([0.0, math.nan, 20.0], [0.0, 0.0, 0.0]),
+  _plan([0.0, math.inf, 20.0], [0.0, 0.0, 0.0]),
+  _plan([0.0, 20.0], [0.0]),
+])
+def test_invalid_plan_data_leaves_curvature_untouched(bad_plan):
+  assert limit_curvature_to_plan(bad_plan, 0.0155, 1.2) == pytest.approx(0.0155)
+
+
+@pytest.mark.parametrize("value", [math.nan, math.inf, -math.inf])
+def test_nonfinite_guard_inputs_disarm(value):
+  assert update_twitch_guard(value, 1.0, False) == 0.0
+  assert update_twitch_guard(TWITCH_GUARD_DURATION, value, False) == 0.0
+
+
+def test_twitch_guard_arms_at_standstill_or_creep_speed():
+  assert update_twitch_guard(0.0, 0.0, True) == TWITCH_GUARD_DURATION
+  assert update_twitch_guard(0.0, 0.3, False) == TWITCH_GUARD_DURATION
+
+
+def test_twitch_guard_decays_after_pullaway_and_expires():
+  remaining = update_twitch_guard(0.0, 0.0, True)
+  remaining = update_twitch_guard(remaining, 1.0, False)
+  assert remaining == pytest.approx(TWITCH_GUARD_DURATION - DT_CTRL)
+
+  for _ in range(int(TWITCH_GUARD_DURATION / DT_CTRL) + 1):
+    remaining = update_twitch_guard(remaining, 1.0, False)
+  assert remaining == 0.0
+
+
+def test_twitch_guard_does_not_arm_while_moving():
+  assert update_twitch_guard(0.0, 1.0, False) == 0.0

selfdrive/modeld/compile_modeld.py

@@ -204,13 +204,14 @@ def _packed_policy_shapes(input_shapes, include_prev_feature=False):
     shapes[key] = tuple(shape)
   if include_prev_feature:
     features_shape = input_shapes["features_buffer"]
-    shapes["prev_feat"] = (features_shape[0], features_shape[2])
+    shapes["prev_feat"] = (features_shape[0], math.prod(features_shape[2:]))
   return shapes, [math.prod(shape) for shape in shapes.values()]
 
 
 def make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip, device):
   queues, npy = make_warp_input_queues(vision_input_shapes, frame_skip, device)
   features_shape = policy_input_shapes["features_buffer"]
+  feature_dim = math.prod(features_shape[2:])
   desire_key = _detect_desire_key(policy_input_shapes)
   desire_shape = policy_input_shapes[desire_key]
   packed_shapes, packed_sizes = _packed_policy_shapes(policy_input_shapes)
@@ -224,7 +225,7 @@ def make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip
   })
   queues.update({
     "feat_q": Tensor(
-      np.zeros((frame_skip * (features_shape[1] - 1) + 1, features_shape[0], features_shape[2]), dtype=np.float32),
+      np.zeros((frame_skip * (features_shape[1] - 1) + 1, features_shape[0], feature_dim), dtype=np.float32),
       device=device,
     ).contiguous().realize(),
     "desire_q": Tensor(
@@ -239,6 +240,7 @@ def make_split_input_queues(vision_input_shapes, policy_input_shapes, frame_skip
 def make_supercombo_input_queues(input_shapes, frame_skip, device):
   queues, npy = make_warp_input_queues(input_shapes, frame_skip, device)
   features_shape = input_shapes["features_buffer"]
+  feature_dim = math.prod(features_shape[2:])
   desire_key = _detect_desire_key(input_shapes)
   desire_shape = input_shapes[desire_key]
   packed_shapes, packed_sizes = _packed_policy_shapes(input_shapes, include_prev_feature=True)
@@ -252,7 +254,7 @@ def make_supercombo_input_queues(input_shapes, frame_skip, device):
   })
   queues.update({
     "feat_q": Tensor(
-      np.zeros((frame_skip * features_shape[1], features_shape[0], features_shape[2]), dtype=np.float32),
+      np.zeros((frame_skip * features_shape[1], features_shape[0], feature_dim), dtype=np.float32),
       device=device,
     ).contiguous().realize(),
     "desire_q": Tensor(
@@ -335,6 +337,7 @@ def make_run_split_policy(vision_runner, policy_runners, metadata, policy_order,
     vision_output = next(iter(vision_runner({road_key: img, wide_key: big_img}).values())).cast("float32")
     new_feature = vision_output[:, vision_features_slice].reshape(1, -1).unsqueeze(0)
     features_buffer = shift_and_sample(feat_q, new_feature, sample_skip_fn)
+    features_buffer = features_buffer.reshape(policy_metadata["input_shapes"]["features_buffer"])
 
     policy_inputs = {
       "features_buffer": features_buffer,
@@ -388,6 +391,7 @@ def make_run_supercombo(model_runner, metadata, frame_skip, image_history_pipeli
     features_buffer = shift_and_sample(
       feat_q, previous_feature.reshape(1, 1, -1), sample_skip_fn,
     )
+    features_buffer = features_buffer.reshape(input_shapes["features_buffer"])
     model_inputs = {
       road_key: img,
       wide_key: big_img,

selfdrive/modeld/modeld.py

@@ -1,4 +1,6 @@
 #!/usr/bin/env python3
+from collections.abc import Callable
+import ctypes
 from functools import cached_property
 import os
 import struct
@@ -159,8 +161,10 @@ class ChestnutState:
     if self.big and "AMD" in Device._opened_devices and self.sends % 100 == 1:
       try:
         smu = Device["AMD"].iface.dev_impl.smu
+        metrics_t = smu.smu_mod.SmuMetricsExternal_t
         smu._send_msg(smu.smu_mod.PPSMC_MSG_TransferTableSmu2Dram, smu.smu_mod.TABLE_SMU_METRICS, timeout=100)
-        metrics = smu.read_table(smu.smu_mod.SmuMetricsExternal_t, smu.smu_mod.TABLE_SMU_METRICS).SmuMetrics
+        metrics_buf = bytearray(smu.adev.vram.view(smu.driver_table_paddr, ctypes.sizeof(metrics_t))[:])
+        metrics = metrics_t.from_buffer(metrics_buf).SmuMetrics
         self.metrics = {
           "tempC": metrics.AvgTemperature[smu.smu_mod.TEMP_HOTSPOT],
           "memoryTempC": metrics.AvgTemperature[smu.smu_mod.TEMP_MEM],
@@ -277,10 +281,11 @@ def _close_tinygrad_disk_cache_connection() -> None:
 def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
                           lat_action_t: float, long_action_t: float, v_ego: float, mlsim: bool,
                           is_v9: bool, is_v14: bool, is_v15: bool, starpilot_toggles,
-                          lat_smooth_seconds=LAT_SMOOTH_SECONDS, long_smooth_seconds=LONG_SMOOTH_SECONDS) -> log.ModelDataV2.Action:
-    if is_v14 or is_v15:
+                          lat_smooth_seconds=LAT_SMOOTH_SECONDS, long_smooth_seconds=LONG_SMOOTH_SECONDS,
+                          is_v16: bool = False) -> log.ModelDataV2.Action:
+    if is_v14 or is_v15 or is_v16:
       desired_curv_unscaled, desired_accel = model_output['action'][0]
-      if is_v15:
+      if is_v15 or is_v16:
         desired_curvature = float(desired_curv_unscaled) / max(1.0, v_ego) ** 2
       else:
         desired_curvature = float(desired_curv_unscaled) / 100.0
@@ -464,6 +469,7 @@ class ModelState:
     self.is_v9 = self.policy_generation == "v9"
     self.is_v14 = self.policy_generation == "v14"
     self.is_v15 = self.policy_generation == "v15"
+    self.is_v16 = self.policy_generation == "v16"
     self.mlsim = is_tinygrad_model_version(self.policy_generation)
     if write_model_version:
       params.put("ModelVersion", self.policy_generation)
@@ -567,7 +573,8 @@ class ModelState:
     self._reset_state()
 
   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:
+          inputs: dict[str, np.ndarray], prepare_only: bool,
+          after_enqueue: Callable[[], None] | None = None) -> dict[str, np.ndarray] | None:
     frames: dict[str, Tensor] = {}
     for key, buf in bufs.items():
       ptr = np.frombuffer(buf.data, dtype=np.uint8).ctypes.data
@@ -613,11 +620,12 @@ class ModelState:
         img=img,
         big_img=big_img,
       )
+    if after_enqueue is not None:
+      after_enqueue()
     outputs = [output.numpy().flatten() for output in output_tensors]
 
     if self.uses_external_gpu and any(not np.isfinite(output).all() for output in outputs):
-      cloudlog.error("external GPU model output not finite, dropping frame")
-      return None
+      raise RuntimeError("external GPU model output not finite")
 
     if self.model_type == "supercombo":
       model_output = outputs[0]
@@ -919,7 +927,17 @@ def main(demo=False):
 
     mt1 = time.perf_counter()
     try:
-      model_output = model.run(bufs, transforms, inputs, prepare_only)
+      send_chestnut = (
+        chestnut_state is not None and
+        run_count % round(ModelConstants.MODEL_FREQ / SERVICE_LIST["chestnutState"].frequency) == 0
+      )
+      model_output = model.run(
+        bufs,
+        transforms,
+        inputs,
+        prepare_only,
+        chestnut_state.send if send_chestnut else None,
+      )
     except Exception:
       if not external_gpu_active or small_model is None:
         raise
@@ -953,7 +971,7 @@ def main(demo=False):
         lat_action_t,
         long_action_t,
         v_ego, model.mlsim, model.is_v9, model.is_v14, model.is_v15, starpilot_toggles,
-        lat_smooth_seconds, long_smooth_seconds,
+        lat_smooth_seconds, long_smooth_seconds, is_v16=model.is_v16,
       )
       prev_action = action
       fill_model_msg(drivingdata_send, modelv2_send, model_output, action,
@@ -982,9 +1000,6 @@ def main(demo=False):
     if sm.updated['starpilotPlan']:
       starpilot_toggles = get_starpilot_toggles(sm)
 
-    if chestnut_state is not None and run_count % round(ModelConstants.MODEL_FREQ / SERVICE_LIST["chestnutState"].frequency) == 0:
-      chestnut_state.send()
-
 if __name__ == "__main__":
   try:
     import argparse

selfdrive/modeld/tests/test_usbgpu_helpers.py

@@ -3,6 +3,7 @@ from types import MethodType
 from types import SimpleNamespace
 
 import numpy as np
+import pytest
 
 from openpilot.selfdrive.modeld import modeld
 from openpilot.selfdrive.modeld.helpers import dump_oob, load_oob, tinygrad_dev_config
@@ -37,17 +38,18 @@ def test_external_gpu_uses_a_longer_load_watchdog():
   assert modeld.BIG_MODEL_RUN_WAIT_TIMEOUT_MS == 3000
 
 
-def test_external_gpu_signal_wait_matches_upstream_busy_poll():
+def test_external_gpu_signal_wait_yields_between_usb_polls(monkeypatch):
   from tinygrad.runtime import ops_amd
 
   sleeps = []
+  monkeypatch.setattr(ops_amd.time, "sleep", sleeps.append)
   signal = ops_amd.AMDSignal.__new__(ops_amd.AMDSignal)
   signal.should_return = False
-  signal.owner = SimpleNamespace(is_usb=lambda: True, iface=SimpleNamespace(sleep=sleeps.append))
+  signal.owner = SimpleNamespace(is_usb=lambda: True, iface=SimpleNamespace(sleep=lambda _: None))
 
   signal._sleep(0)
 
-  assert sleeps == []
+  assert sleeps == [ops_amd.AMD_USB_POLL_US / 1e6]
 
 
 def test_native_amd_signal_keeps_existing_short_wait_behavior():
@@ -201,7 +203,7 @@ def test_external_gpu_load_finishes_before_native_model_can_start(monkeypatch):
   ]
 
 
-def test_external_gpu_nonfinite_outputs_are_dropped_without_escalating(monkeypatch):
+def test_external_gpu_nonfinite_outputs_trigger_fallback(monkeypatch):
   class FakeTensor:
     @staticmethod
     def from_blob(*_args, **_kwargs):
@@ -235,13 +237,7 @@ def test_external_gpu_nonfinite_outputs_are_dropped_without_escalating(monkeypat
   state.image_history_pipeline = modeld.IMAGE_HISTORY_IN_POLICY
   state.warp_enqueue = lambda **_kwargs: object()
   state.run_policy = lambda **_kwargs: (FakeOutput(),)
-  state._reset_state = MethodType(
-    lambda self: (_ for _ in ()).throw(AssertionError("upstream does not reset or escalate transient non-finite output")),
-    state,
-  )
-
   monkeypatch.setattr(modeld, "Tensor", FakeTensor)
-  monkeypatch.setattr(modeld.cloudlog, "error", lambda *_args, **_kwargs: None)
   buffers = {
     "img": SimpleNamespace(data=bytearray(4)),
     "big_img": SimpleNamespace(data=bytearray(4)),
@@ -252,8 +248,10 @@ def test_external_gpu_nonfinite_outputs_are_dropped_without_escalating(monkeypat
   }
   inputs = {"desire_pulse": np.zeros(8, dtype=np.float32)}
 
-  for _ in range(10):
-    assert state.run(buffers, transforms, inputs, False) is None
+  callbacks = []
+  with pytest.raises(RuntimeError, match="external GPU model output not finite"):
+    state.run(buffers, transforms, inputs, False, lambda: callbacks.append("sent"))
+  assert callbacks == ["sent"]
 
 
 def test_out_of_band_artifact_round_trip():

selfdrive/ui/mici/layouts/settings/software.py

@@ -1,6 +1,11 @@
+import json
 import os
 import threading
+import time
+import urllib.error
+import urllib.request
 from collections.abc import Callable
+from dataclasses import dataclass
 from enum import IntEnum
 
 import pyray as rl
@@ -18,6 +23,40 @@ from openpilot.system.ui.widgets.label import UnifiedLabel
 from openpilot.system.ui.widgets.scroller import NavScroller
 
 UPDATER_TIMEOUT = 10.0
+FAST_UPDATE_HOLD_SECONDS = 0.8
+FAST_UPDATE_POLL_SECONDS = 0.5
+FAST_UPDATE_REQUEST_TIMEOUT = 5.0
+
+
+@dataclass
+class FastUpdateDisplayState:
+  stage: str = "idle"
+  message: str = ""
+  error: str = ""
+
+
+def _galaxy_api_url(path: str) -> str:
+  port = os.getenv("SP_GALAXY_PORT", "8082")
+  return f"http://127.0.0.1:{port}/api/update/fast{path}"
+
+
+def _galaxy_fast_update_request(path: str = "", method: str = "GET") -> dict:
+  request = urllib.request.Request(_galaxy_api_url(path), method=method)
+  try:
+    with urllib.request.urlopen(request, timeout=FAST_UPDATE_REQUEST_TIMEOUT) as response:
+      payload = json.loads(response.read().decode("utf-8"))
+  except urllib.error.HTTPError as error:
+    try:
+      payload = json.loads(error.read().decode("utf-8"))
+    except (json.JSONDecodeError, UnicodeDecodeError):
+      payload = {}
+    raise RuntimeError(payload.get("error") or str(error)) from error
+  except (OSError, urllib.error.URLError, json.JSONDecodeError, UnicodeDecodeError) as error:
+    raise RuntimeError(f"Galaxy fast update unavailable: {error}") from error
+
+  if not isinstance(payload, dict):
+    raise RuntimeError("Galaxy returned an invalid fast update response")
+  return payload
 
 
 def _split_description(desc: str) -> tuple[str, str, str, str] | None:
@@ -101,6 +140,10 @@ class CheckUpdateButton(BigButton):
     self._waiting_for_updater_t: float | None = None
     self._hide_value_t: float | None = None
     self._state: UpdaterState = UpdaterState.IDLE
+    self._press_start_t: float | None = None
+    self._press_action = ""
+    self._fast_update_state = FastUpdateDisplayState()
+    self._fast_update_state_lock = threading.Lock()
 
     ui_state.add_offroad_transition_callback(self.offroad_transition)
 
@@ -108,9 +151,24 @@ class CheckUpdateButton(BigButton):
     if ui_state.is_offroad():
       self.set_enabled(True)
 
+  def _handle_mouse_press(self, mouse_pos: MousePos):
+    super()._handle_mouse_press(mouse_pos)
+    self._press_start_t = rl.get_time()
+    self._press_action = self.get_value()
+
   def _handle_mouse_release(self, mouse_pos: MousePos):
+    held_for = 0.0 if self._press_start_t is None else rl.get_time() - self._press_start_t
+    self._press_start_t = None
     super()._handle_mouse_release(mouse_pos)
 
+    if held_for >= FAST_UPDATE_HOLD_SECONDS:
+      self._show_fast_update_confirmation()
+      return
+
+    if self._get_fast_update_state().stage == "error":
+      self._set_fast_update_state(stage="idle")
+      return
+
     if not system_time_valid():
       dlg = BigDialog("", tr("Please connect to Wi-Fi to update."))
       gui_app.push_widget(dlg)
@@ -121,13 +179,75 @@ class CheckUpdateButton(BigButton):
     self.set_icon(self._txt_update_icon)
 
     def run():
-      if self.get_value() == "download update":
+      if self._press_action == "download update":
         _request_update_download()
       else:
         _request_update_check()
 
     threading.Thread(target=run, daemon=True).start()
 
+  def _show_fast_update_confirmation(self):
+    if not system_time_valid():
+      gui_app.push_widget(BigDialog("", tr("Please connect to Wi-Fi to update.")))
+      return
+    if ui_state.started:
+      return
+
+    gui_app.push_widget(BigConfirmationDialog(
+      "slide to\nfast update",
+      self._txt_update_icon,
+      self._start_fast_update,
+      red=True,
+    ))
+
+  def _set_fast_update_state(self, *, stage: str, message: str = "", error: str = ""):
+    with self._fast_update_state_lock:
+      self._fast_update_state = FastUpdateDisplayState(stage, message, error)
+
+  def _get_fast_update_state(self) -> FastUpdateDisplayState:
+    with self._fast_update_state_lock:
+      state = self._fast_update_state
+      return FastUpdateDisplayState(state.stage, state.message, state.error)
+
+  def _start_fast_update(self):
+    if ui_state.started:
+      return
+
+    state = self._get_fast_update_state()
+    if state.stage not in ("idle", "error"):
+      return
+
+    self._set_fast_update_state(stage="starting", message="starting fast update...")
+    threading.Thread(target=self._run_fast_update, daemon=True).start()
+
+  def _run_fast_update(self):
+    try:
+      _galaxy_fast_update_request(method="POST")
+
+      while True:
+        status = _galaxy_fast_update_request("/status")
+        stage = str(status.get("stage") or "updating")
+        error = str(status.get("lastError") or "").strip()
+        message = str(
+          status.get("progressDetail") or
+          status.get("progressLabel") or
+          status.get("message") or
+          "fast update in progress..."
+        ).strip()
+
+        if error or stage == "error":
+          self._set_fast_update_state(stage="error", message="fast update failed", error=error or message)
+          return
+
+        self._set_fast_update_state(stage=stage, message=message)
+        if not bool(status.get("running")):
+          return
+        time.sleep(FAST_UPDATE_POLL_SECONDS)
+    except Exception as error:
+      current = self._get_fast_update_state()
+      if current.stage != "rebooting":
+        self._set_fast_update_state(stage="error", message="fast update failed", error=str(error))
+
   def set_value(self, value: str):
     super().set_value(value)
     self.set_text("" if value else "check for update")
@@ -136,9 +256,25 @@ class CheckUpdateButton(BigButton):
     super()._update_state()
 
     if ui_state.started:
+      self._press_start_t = None
       self.set_enabled(False)
       return
 
+    fast_update_state = self._get_fast_update_state()
+    if fast_update_state.stage != "idle":
+      self.set_rotate_icon(fast_update_state.stage not in ("error", "rebooting"))
+      if fast_update_state.stage == "error":
+        self.set_enabled(True)
+        self.set_value(fast_update_state.error or fast_update_state.message)
+      elif fast_update_state.stage == "rebooting":
+        self.set_enabled(False)
+        self.set_value("update complete\nrebooting...")
+      else:
+        self.set_enabled(False)
+        self.set_value(fast_update_state.message or "fast update in progress...")
+      self.set_text("fast update")
+      return
+
     updater_state = ui_state.params.get("UpdaterState") or ""
 
     if self._state == UpdaterState.WAITING_FOR_UPDATER:

selfdrive/ui/mici/layouts/settings/tests/test_software_fast_update.py

@@ -0,0 +1,98 @@
+from types import SimpleNamespace
+
+from openpilot.selfdrive.ui.mici.layouts.settings import software
+from openpilot.selfdrive.ui.mici.widgets.button import BigButton
+
+
+def _make_button() -> software.CheckUpdateButton:
+  button = object.__new__(software.CheckUpdateButton)
+  button._press_start_t = None
+  button._press_action = ""
+  button._fast_update_state = software.FastUpdateDisplayState()
+  button._fast_update_state_lock = software.threading.Lock()
+  return button
+
+
+def test_fast_update_uses_local_galaxy_api(monkeypatch):
+  monkeypatch.delenv("SP_GALAXY_PORT", raising=False)
+
+  assert software._galaxy_api_url("") == "http://127.0.0.1:8082/api/update/fast"
+  assert software._galaxy_api_url("/status") == "http://127.0.0.1:8082/api/update/fast/status"
+
+
+def test_long_press_opens_fast_update_confirmation(monkeypatch):
+  button = _make_button()
+  button._press_start_t = 1.0
+  confirmation_calls = []
+
+  monkeypatch.setattr(software.rl, "get_time", lambda: 1.0 + software.FAST_UPDATE_HOLD_SECONDS)
+  monkeypatch.setattr(BigButton, "_handle_mouse_release", lambda *_args: None)
+  monkeypatch.setattr(button, "_show_fast_update_confirmation", lambda: confirmation_calls.append(True))
+
+  button._handle_mouse_release(SimpleNamespace())
+
+  assert confirmation_calls == [True]
+
+
+def test_short_press_keeps_normal_updater_path(monkeypatch):
+  button = _make_button()
+  button._press_start_t = 1.0
+  button._press_action = "download update"
+  downloads = []
+  confirmations = []
+
+  monkeypatch.setattr(software.rl, "get_time", lambda: 1.0 + software.FAST_UPDATE_HOLD_SECONDS - 0.1)
+  monkeypatch.setattr(BigButton, "_handle_mouse_release", lambda *_args: None)
+  monkeypatch.setattr(software, "system_time_valid", lambda: True)
+  monkeypatch.setattr(software, "_request_update_download", lambda: downloads.append(True))
+  monkeypatch.setattr(button, "_show_fast_update_confirmation", lambda: confirmations.append(True))
+
+  class ImmediateThread:
+    def __init__(self, target, daemon):
+      self.target = target
+
+    def start(self):
+      self.target()
+
+  monkeypatch.setattr(software.threading, "Thread", ImmediateThread)
+  button.set_enabled = lambda *_args: None
+  button.set_icon = lambda *_args: None
+  button._state = software.UpdaterState.IDLE
+  button._txt_update_icon = None
+
+  button._handle_mouse_release(SimpleNamespace())
+
+  assert downloads == [True]
+  assert confirmations == []
+
+
+def test_fast_update_worker_uses_galaxy_endpoint(monkeypatch):
+  button = _make_button()
+  requests = []
+  responses = [
+    {"message": "started"},
+    {
+      "running": True,
+      "stage": "updating",
+      "progressDetail": "Fetching latest shallow commit...",
+      "lastError": "",
+    },
+    {
+      "running": False,
+      "stage": "rebooting",
+      "progressDetail": "Update complete. Please wait for device to reboot.",
+      "lastError": "",
+    },
+  ]
+
+  def request(path="", method="GET"):
+    requests.append((path, method))
+    return responses.pop(0)
+
+  monkeypatch.setattr(software, "_galaxy_fast_update_request", request)
+  monkeypatch.setattr(software.time, "sleep", lambda *_args: None)
+
+  button._run_fast_update()
+
+  assert requests == [("", "POST"), ("/status", "GET"), ("/status", "GET")]
+  assert button._get_fast_update_state().stage == "rebooting"

selfdrive/ui/mici/onroad/hud_renderer.py

@@ -158,7 +158,7 @@ class HudRenderer(Widget):
     self._txt_exclamation_point: rl.Texture = gui_app.texture('icons_mici/exclamation_point.png', 44, 44)
     self._txt_egpu_loading: rl.Texture = gui_app.texture('icons_mici/egpu_loading.png', 60, 44)
     self._txt_egpu_green: rl.Texture = gui_app.texture('icons_mici/egpu_green.png', 60, 44)
-    self._txt_egpu_orange: rl.Texture = gui_app.texture('icons_mici/egpu_orange.png', 60, 44)
+    self._txt_egpu_orange: rl.Texture = gui_app.texture('icons_mici/egpu_orange.png', 75, 44)
     self._txt_egpu_crossed: rl.Texture = gui_app.texture('icons_mici/egpu_crossed.png', 60, 52)
     self._egpu_icon: rl.Texture | None = None
 

selfdrive/ui/mici/onroad/model_renderer.py

@@ -6,6 +6,7 @@ from cereal import messaging, car
 from dataclasses import dataclass, field
 from openpilot.common.constants import CV
 from openpilot.common.filter_simple import FirstOrderFilter
+from openpilot.selfdrive.controls.lib.lane_centering import get_lane_centering_visual_direction
 from openpilot.selfdrive.locationd.calibrationd import HEIGHT_INIT
 from openpilot.selfdrive.ui.lib.starpilot_theme import get_param_color, get_theme_color, get_visual_color, is_stock_color_scheme, with_alpha
 from openpilot.selfdrive.ui.onroad.starpilot.rainbow_path import RainbowPath
@@ -361,7 +362,31 @@ class ModelRenderer(Widget):
 
     return LeadVehicle(glow=glow, chevron=chevron, fill_alpha=int(fill_alpha))
 
-  def _lane_line_palette(self) -> tuple[bool, rl.Color, rl.Color, bool]:
+  def _lane_centering_direction(self) -> int:
+    toggles = ui_state.starpilot_toggles
+    sm = ui_state.sm
+    if not sm.valid.get("modelV2", False) or not sm.valid.get("carState", False):
+      return 0
+
+    car_state = sm["carState"]
+    applied_correction = None
+    if sm.valid.get("controlsState", False):
+      try:
+        applied_correction = sm["controlsState"].desiredCurvature - sm["modelV2"].action.desiredCurvature
+      except (AttributeError, TypeError, ValueError):
+        pass
+    return get_lane_centering_visual_direction(
+      sm["modelV2"], car_state.vEgo,
+      toggles.get("lane_center_offset", 0.0),
+      toggles.get("lane_centering_e2e_authority", 1.0),
+      bool(toggles.get("lane_centering", False)),
+      ui_state.status == UIStatus.ENGAGED or ui_state.always_on_lateral_active,
+      bool(toggles.get("lane_centering_pause_on_signal", True)),
+      bool(car_state.leftBlinker or car_state.rightBlinker),
+      applied_correction,
+    )
+
+  def _lane_line_palette(self) -> tuple[bool, rl.Color, rl.Color, int]:
     stock_scheme = is_stock_color_scheme(self._params)
     line_status = UIStatus.ENGAGED if ui_state.status == UIStatus.DISENGAGED and ui_state.always_on_lateral_active else ui_state.status
 
@@ -374,15 +399,15 @@ class ModelRenderer(Widget):
     if lane_color is None:
       lane_color = STOCK_LANE_LINES_COLOR if stock_scheme else get_theme_color("LaneLines", STOCK_LANE_LINES_COLOR)
 
-    lane_centering_active = bool(ui_state.starpilot_toggles.get("lane_centering", False)) and (
-      ui_state.status == UIStatus.ENGAGED or ui_state.always_on_lateral_active
-    )
-    return stock_scheme, edge_color, lane_color, lane_centering_active
+    lane_centering_direction = self._lane_centering_direction()
+    return stock_scheme, edge_color, lane_color, lane_centering_direction
 
   def _get_ll_color(self, prob: float, adjacent: bool, left: bool, stock_scheme: bool,
-                    edge_color: rl.Color, lane_color: rl.Color, lane_centering_active: bool = False):
+                    edge_color: rl.Color, lane_color: rl.Color, lane_centering_direction: int = 0):
     alpha = np.clip(prob, 0.0, 0.7)
-    if lane_centering_active:
+    lane_centering_line = adjacent and ((lane_centering_direction > 0 and not left) or
+                                        (lane_centering_direction < 0 and left))
+    if lane_centering_line:
       color = rl.Color(OCEAN_BLUE_LANE_LINES_COLOR.r, OCEAN_BLUE_LANE_LINES_COLOR.g,
                        OCEAN_BLUE_LANE_LINES_COLOR.b, int(alpha * OCEAN_BLUE_LANE_LINES_COLOR.a))
     elif adjacent:
@@ -408,13 +433,13 @@ class ModelRenderer(Widget):
   def _draw_lane_lines(self):
     """Draw lane lines and road edges"""
     """Two closest lines should be green (lane line or road edges)"""
-    stock_scheme, edge_color, lane_color, lane_centering_active = self._lane_line_palette()
+    stock_scheme, edge_color, lane_color, lane_centering_direction = self._lane_line_palette()
     for i, lane_line in enumerate(self._lane_lines):
       if lane_line.projected_points.size == 0:
         continue
 
       color = self._get_ll_color(float(self._lane_line_probs[i]), i in (1, 2), i in (0, 1),
-                                 stock_scheme, edge_color, lane_color, lane_centering_active)
+                                 stock_scheme, edge_color, lane_color, lane_centering_direction)
       draw_polygon(self._rect, lane_line.projected_points, color)
 
     for i, road_edge in enumerate(self._road_edges):

selfdrive/ui/onroad/model_renderer.py

@@ -5,6 +5,7 @@ from cereal import messaging, car
 from dataclasses import dataclass, field
 from openpilot.common.filter_simple import FirstOrderFilter
 from openpilot.common.constants import CV
+from openpilot.selfdrive.controls.lib.lane_centering import get_lane_centering_visual_direction
 from openpilot.selfdrive.locationd.calibrationd import HEIGHT_INIT
 from openpilot.selfdrive.ui.lib.starpilot_theme import get_param_color, get_theme_color, get_visual_color, is_stock_color_scheme, with_alpha
 from openpilot.selfdrive.ui.onroad.radar_tracks import project_radar_points
@@ -369,15 +370,35 @@ class ModelRenderer(Widget):
 
     return LeadVehicle(glow=glow, chevron=chevron, fill_alpha=int(fill_alpha))
 
+  def _lane_centering_direction(self) -> int:
+    toggles = ui_state.starpilot_toggles
+    sm = ui_state.sm
+    if not sm.valid.get("modelV2", False) or not sm.valid.get("carState", False):
+      return 0
+
+    car_state = sm["carState"]
+    applied_correction = None
+    if sm.valid.get("controlsState", False):
+      try:
+        applied_correction = sm["controlsState"].desiredCurvature - sm["modelV2"].action.desiredCurvature
+      except (AttributeError, TypeError, ValueError):
+        pass
+    return get_lane_centering_visual_direction(
+      sm["modelV2"], car_state.vEgo,
+      toggles.get("lane_center_offset", 0.0),
+      toggles.get("lane_centering_e2e_authority", 1.0),
+      bool(toggles.get("lane_centering", False)),
+      ui_state.status == UIStatus.ENGAGED or ui_state.always_on_lateral_active,
+      bool(toggles.get("lane_centering_pause_on_signal", True)),
+      bool(car_state.leftBlinker or car_state.rightBlinker),
+      applied_correction,
+    )
+
   def _draw_lane_lines(self):
     """Draw lane lines and road edges"""
-    lane_centering_active = bool(ui_state.starpilot_toggles.get("lane_centering", False)) and (
-      ui_state.status == UIStatus.ENGAGED or ui_state.always_on_lateral_active
-    )
+    lane_centering_direction = self._lane_centering_direction()
     lane_lines_override = get_param_color(self._params, "LaneLinesColor", STOCK_LANE_LINES_COLOR.a)
-    if lane_centering_active:
-      lane_lines_color = OCEAN_BLUE_LANE_LINES_COLOR
-    elif lane_lines_override is not None:
+    if lane_lines_override is not None:
       lane_lines_color = lane_lines_override
     elif is_stock_color_scheme(self._params):
       lane_lines_color = STOCK_LANE_LINES_COLOR
@@ -389,7 +410,10 @@ class ModelRenderer(Widget):
         continue
 
       alpha = np.clip(self._lane_line_probs[i], 0.0, 0.7)
-      color = with_alpha(lane_lines_color, int(alpha * lane_lines_color.a))
+      lane_centering_line = (lane_centering_direction > 0 and i == 2) or \
+                            (lane_centering_direction < 0 and i == 1)
+      line_color = OCEAN_BLUE_LANE_LINES_COLOR if lane_centering_line else lane_lines_color
+      color = with_alpha(line_color, int(alpha * line_color.a))
       draw_polygon(self._rect, lane_line.projected_points, color)
 
     for i, road_edge in enumerate(self._road_edges):

starpilot/common/accel_profile.py

@@ -3,8 +3,6 @@ from __future__ import annotations
 
 import math
 
-from openpilot.selfdrive.controls.lib.longitudinal_planner import get_max_accel
-
 ACCELERATION_PROFILES = {
   "STANDARD": 0,
   "ECO": 1,

starpilot/controls/lib/starpilot_vcruise.py

@@ -19,7 +19,9 @@ CSC_MIN_SPEED = CITY_SPEED_LIMIT * CV.MPH_TO_MS
 CSC_CURVE_RELEASE_HOLD_TIME = 0.75
 OVERRIDE_FORCE_STOP_TIMER = 10
 STANDSTILL_FORCE_STOP_CLEAR_TIME = 0.75
-STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME = 5.0
+# Open-loop — green is undetectable at standstill, so this only needs to cover the
+# handoff to CEM+model ownership. Extra seconds are pure departure lag.
+STANDSTILL_FORCE_STOP_LIGHT_HOLD_TIME = 2.0
 FORCE_STOP_LIGHT_CLEAR_TIME = 0.5
 SLC_LEAD_DROP_RELAXATION_MIN_SPEED = 20.0 * CV.MPH_TO_MS
 SLC_LEAD_DROP_RELAXATION_MIN_DISTANCE = 30.0
@@ -57,6 +59,8 @@ LEAD_VETO_M_OVERRIDES = {
 }
 FORCE_STOP_APPROACH_DECEL = 0.65  # m/s^2 — speed ceiling before commit. LOWER = more early
                           # braking; don't go under FORCE_STOP_MODEL_APPROACH_DECEL
+# approachStopLength is published RAW: model_length converges from above, so rate-limiting
+# it inward freezes it far out and the constraint never binds. Tried, measured, don't re-add.
 ADAS_MAX_MS = 17.88       # 40 mph — cross-street ADAS guard
 DASH_SEED_M = 27.0        # ~88 ft — typical ADAS detection distance, used to snap
                           # tracked length closer when dashboard confirms a sign
@@ -74,6 +78,14 @@ FORCE_STOP_TURN_VETO_STEERING_ANGLE = 25.0
 FORCE_STOP_CURVE_VETO_MAX_ROAD_CURVATURE = 0.003
 FORCE_STOP_TURN_VETO_STOP_SEEN_HOLD_TIME = 4.0
 FORCE_STOP_DISTANCE_REANCHOR_MIN_GAP = 3.0  # m — ignore small model-horizon noise
+FORCE_STOP_REANCHOR_MIN_M = 40.0  # m — inside this only ratchet down; shouldStop doesn't
+                          # assert until ~10 m, so horizon jitter would release the stop
+FORCE_STOP_CAP_SLACK_M = 15.0  # m — the line can't move away, so tracked can never exceed
+                          # what it was at commit minus distance driven. Slack covers an
+                          # under-read at commit; without it that would stop us short.
+FORCE_STOP_CAP_TAPER_M = 60.0  # m — slack fades to 0 as the cap closes. The solver aims at
+                          # tracked, so slack held near the line is braking for a stop bar
+                          # that far past the real one.
 
 # Knob bounds (mirror of UI slider; defense in depth)
 OFFSET_FT_MIN = -20
@@ -177,9 +189,11 @@ class StarPilotVCruise:
     self.force_stop_from_light = False
     self.force_stop_light_clear_since = None
     self.controls_enabled_previously = False
+    self.approach_stop_length = 0.0  # published as starpilotPlan.approachStopLength
     # Kinematic distance estimator. Same attribute also published as
     # starpilotPlan.forcingStopLength, so the existing reader keeps working.
     self.tracked_model_length = 0.0
+    self.force_stop_distance_cap = 0.0  # odometry ceiling, re-seeded until commit
 
     self.stop_sign_confirmed = False
     self.stop_seen_on_approach_at = None
@@ -603,6 +617,9 @@ class StarPilotVCruise:
     offset_ft = max(OFFSET_FT_MIN, min(OFFSET_FT_MAX, offset_ft_raw))
     offset_m = offset_ft * FT_TO_M
 
+    # cleared on every path; only the far-approach envelope below republishes it
+    self.approach_stop_length = 0.0
+
     if force_standstill_enabled and not self.override_force_standstill:
       self.forcing_stop = True
       self.tracked_model_length = 0.0
@@ -630,7 +647,7 @@ class StarPilotVCruise:
           model_wants_stop = False
         if (
           not dash_active and
-          self.tracked_model_length > force_stop_handoff_m and
+          self.tracked_model_length > max(force_stop_handoff_m, FORCE_STOP_REANCHOR_MIN_M) and
           not model_wants_stop and
           model_length > self.tracked_model_length + FORCE_STOP_DISTANCE_REANCHOR_MIN_GAP and
           (
@@ -642,6 +659,12 @@ class StarPilotVCruise:
           self.tracked_model_length = model_length
         else:
           self.tracked_model_length = min(self.tracked_model_length, model_length)
+        # Odometry ceiling: the line can't recede, so a re-anchor may never exceed what we
+        # had at commit minus what we've driven. Bounds a ballooning horizon (seen +95 m)
+        # that the REANCHOR_MIN floor can't catch, since that floor trusts the estimate.
+        self.force_stop_distance_cap = max(self.force_stop_distance_cap - (v_ego * DT_MDL), 0.0)
+        cap_slack = FORCE_STOP_CAP_SLACK_M * min(self.force_stop_distance_cap / FORCE_STOP_CAP_TAPER_M, 1.0)
+        self.tracked_model_length = min(self.tracked_model_length, self.force_stop_distance_cap + cap_slack)
         if dash_active:
           if model_length < DASH_MODEL_AGREE_M:
             self.tracked_model_length = min(self.tracked_model_length, DASH_SEED_M)
@@ -675,6 +698,7 @@ class StarPilotVCruise:
       self.stop_sign_confirmed = False
 
       self.tracked_model_length = self.starpilot_planner.model_length
+      self.force_stop_distance_cap = self.tracked_model_length
 
       targets = [v_cruise]
       if self.csc_target >= CSC_MIN_SPEED:
@@ -720,6 +744,8 @@ class StarPilotVCruise:
         adjacent_stop_d = self._get_adjacent_stop_distance(sm)
         if adjacent_stop_d is not None:
           approach_d = min(approach_d, adjacent_stop_d)
+        # pre-offset, so it hands off to forcingStopLength at commit without a step
+        self.approach_stop_length = max(approach_d, 0.0)
         approach_d += offset_m + force_stop_distance_bias_m
         if approach_d > force_stop_handoff_m:
           targets.append(math.sqrt(2.0 * FORCE_STOP_APPROACH_DECEL * (approach_d - force_stop_handoff_m)))

starpilot/controls/starpilot_planner.py

@@ -31,7 +31,10 @@ from openpilot.starpilot.controls.lib.starpilot_vcruise import StarPilotVCruise
 from openpilot.starpilot.controls.lib.weather_checker import WeatherChecker
 
 RADARLESS_TRACK_HOLD_TIME = 0.45
-FORCE_STOP_JERK_SCALE = 0.32  # accel-change cost multiplier while forcing_stop (125 -> ~40)
+FORCE_STOP_JERK_SCALE = 0.20  # accel-change cost multiplier for the whole stop approach,
+                              # envelope included (125 -> 25). Lower = reaches the braking
+                              # target sooner; it does not make the target deeper. Response
+                              # is super-linear here, so raise it if onset feels like a step.
 FORCE_STOP_JERK_SCALE_OVERRIDES = {
   # The Elantra's current force-stop ramp is smooth, but it waits too long
   # before building decel and then arrives at the initial brake too abruptly.
@@ -308,7 +311,9 @@ class StarPilotPlanner:
     except (KeyError, IndexError, TypeError, AttributeError):
       car_params = None
 
-    if self.starpilot_vcruise.forcing_stop:
+    # Also while the far-approach envelope is running: at onset the ramp reaches only
+    # ~-0.5 m/s^2 after a second, so the first seconds of a detected red are mostly lost.
+    if self.starpilot_vcruise.forcing_stop or self.starpilot_vcruise.approach_stop_length > 0.0:
       jerk_scale = get_force_stop_jerk_scale(car_params)
     elif self.tracking_lead:
       # Elantra vision leads can hand off from cruise to lead0 while closing
@@ -346,6 +351,7 @@ class StarPilotPlanner:
 
     starpilotPlan.forcingStop = self.starpilot_vcruise.forcing_stop
     starpilotPlan.forcingStopLength = self.starpilot_vcruise.tracked_model_length
+    starpilotPlan.approachStopLength = float(self.starpilot_vcruise.approach_stop_length)
     starpilotPlan.stopSignConfirmed = self.starpilot_vcruise.stop_sign_confirmed
 
     starpilotPlan.starpilotEvents = self.starpilot_events.events.to_msg()

tinygrad_repo/tinygrad/runtime/ops_amd.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 from typing import cast
-import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit
+import os, ctypes, struct, hashlib, functools, importlib, mmap, errno, array, contextlib, sys, weakref, itertools, collections, atexit, time
 assert sys.platform != 'win32'
 from dataclasses import dataclass
 from tinygrad.runtime.support.hcq import HCQCompiled, HCQAllocator, HCQBuffer, HWQueue, CLikeArgsState, HCQSignal, HCQProgram, FileIOInterface
@@ -25,6 +25,7 @@ SQTT = ContextVar("SQTT", abs(VIZ.value)>=2)
 SQTT_ITRACE_SE_MASK, SQTT_LIMIT_SE, SQTT_SIMD_SEL, SQTT_TOKEN_EXCLUDE = \
   ContextVar("SQTT_ITRACE_SE_MASK", 0b11), ContextVar("SQTT_LIMIT_SE", 0), ContextVar("SQTT_SIMD_SEL", 0), ContextVar("SQTT_TOKEN_EXCLUDE", 0)
 PMC = ContextVar("PMC", abs(VIZ.value)>=2)
+AMD_USB_POLL_US = getenv("AMD_USB_POLL_US", 500) # microseconds to sleep between USB signal polls. 0 disables
 EVENT_INDEX_PARTIAL_FLUSH = 4 # based on a comment in nvd.h
 WAIT_REG_MEM_FUNCTION_EQ  = 3 # ==
 WAIT_REG_MEM_FUNCTION_NEQ = 4 # !=
@@ -45,8 +46,9 @@ class AMDSignal(HCQSignal):
   def __init__(self, *args, **kwargs): super().__init__(*args, **{**kwargs, 'timestamp_divider': 100})
 
   def _sleep(self, time_spent_since_last_sleep_ms:int):
-    # Reasonable to sleep for long workloads (which take more than 200ms) and only timeline signals.
-    if time_spent_since_last_sleep_ms > 200 and self.owner is not None: self.owner.iface.sleep(200)
+    # USB signals live in VRAM across the link, so yield between polls. Native AMD only blocks after 200 ms.
+    if self.owner is not None and self.owner.is_usb() and AMD_USB_POLL_US: time.sleep(AMD_USB_POLL_US / 1e6)
+    elif time_spent_since_last_sleep_ms > 200 and self.owner is not None: self.owner.iface.sleep(200)
 
 class AMDComputeQueue(HWQueue):
   def __init__(self, dev:AMDDevice):