models: retain SNPE runner support (#522)

* tinygrad with snpe * force with snpe to validate * fix path * fix more paths * Adjust modeld execution logic based on active model runner Introduced a check to conditionally execute `modeld` based on the active model runner. Added support for distinguishing between SNPE and TinyGrad runners using new helper functions and updated `custom.capnp` definitions. This change optimizes process management by ensuring compatibility with the selected model runner. * Refactor modeld process function checks. Introduce `is_stock_model` to clarify logic and replace direct uses of `is_snpe_model` where the stock model condition is needed. Additionally, rename the duplicate "modeld" process in sunnyPilot to "modeld_snpe" for clarity and consistency. * ignore tg * fix process name * ruff * fix thneed paths * mypy * remove our own * use upstream compile3 * fix thneed * try this * Revert "remove our own" This reverts commit 1cf4f57502. * try using compile2.py again * add back symlink * fix path * more fix * wrong path again * Revert "wrong path again" This reverts commit f5301c19d5. * update * hardcode path to our submodule * force path * try this * fix file name * try this * again * Revert "again" This reverts commit 17c8cd7376. * Revert "try this" This reverts commit 767f78bbcf. * Revert "fix file name" This reverts commit 485eef68da. * Revert "try this" This reverts commit 41fef87680. * Revert "force path" This reverts commit 5c3b408937. * Revert "hardcode path to our submodule" This reverts commit 5ee1950b6f. * Revert "update" This reverts commit fb313bd7fb. * Reapply "wrong path again" This reverts commit 309639aeb3. * Revert "wrong path again" This reverts commit f5301c19d5. * Revert "more fix" This reverts commit 23dd423e78. * Revert "fix path" This reverts commit 75d338f2bd. * Revert "add back symlink" This reverts commit 9f71ad0b8a. * Revert "try using compile2.py again" This reverts commit 914117d2e1. * Reapply "remove our own" This reverts commit b1996377b3. * don't even compile anymore * need it for default snpe model * add to lfs * bring onnx back for sim * must add this back * need this --------- Co-authored-by: DevTekVE <devtekve@gmail.com>
2026-06-08 11:25:51 +08:00 · 2025-01-05 17:47:27 -05:00
parent c12c22eac7
commit f84d27c4ee
56 changed files with 2879 additions and 4 deletions
--- a/.gitattributes
+++ b/.gitattributes
@@ -3,6 +3,8 @@
 # to move existing files into LFS:
 # git add --renormalize .
 *.onnx filter=lfs diff=lfs merge=lfs -text
+*.thneed filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
 *.svg filter=lfs diff=lfs merge=lfs -text
 *.png filter=lfs diff=lfs merge=lfs -text
 *.gif filter=lfs diff=lfs merge=lfs -text
--- a/.gitignore
+++ b/.gitignore
@@ -74,6 +74,7 @@ comma*.sh
 selfdrive/modeld/thneed/compile
 selfdrive/modeld/models/*.thneed
 selfdrive/modeld/models/*.pkl
+sunnypilot/modeld/thneed/compile

 *.bz2
 *.zst
--- a/.gitmodules
+++ b/.gitmodules
@@ -15,4 +15,4 @@
  url = https://github.com/commaai/teleoprtc
 [submodule "tinygrad"]
  path = tinygrad_repo
-  url = https://github.com/tinygrad/tinygrad.git
+  url = https://github.com/commaai/tinygrad.git
--- a/2
+++ b/2
@@ -396,6 +396,8 @@ SConscript(['third_party/SConscript'])

 SConscript(['selfdrive/SConscript'])

+SConscript(['sunnypilot/SConscript'])
+
 if Dir('#tools/cabana/').exists() and GetOption('extras'):
  SConscript(['tools/replay/SConscript'])
  if arch != "larch64":
--- a/cereal/custom.capnp
+++ b/cereal/custom.capnp
@@ -64,6 +64,11 @@ struct ModelManagerSP @0xaedffd8f31e7b55d {
    progress @1 :Float32;
    eta @2 :UInt32;
  }
+  
+  enum Runner {
+    snpe @0;
+    tinygrad @1;
+  }

  struct ModelBundle {
    index @0 :UInt32;
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -137,7 +137,7 @@ allow-direct-references = true

 [tool.pytest.ini_options]
 minversion = "6.0"
-addopts = "--ignore=openpilot/ --ignore=opendbc/ --ignore=panda/ --ignore=rednose_repo/ --ignore=tinygrad_repo/ --ignore=teleoprtc_repo/ --ignore=msgq/  -Werror --strict-config --strict-markers --durations=10 -n auto --dist=loadgroup"
+addopts = "--ignore=openpilot/ --ignore=opendbc/ --ignore=panda/ --ignore=rednose_repo/ --ignore=tinygrad_repo/ --ignore=teleoprtc_repo/ --ignore=msgq/ --ignore=sunnypilot/tinygrad_repo/  -Werror --strict-config --strict-markers --durations=10 -n auto --dist=loadgroup"
 cpp_files = "test_*"
 cpp_harness = "selfdrive/test/cpp_harness.py"
 python_files = "test_*.py"
--- a/sunnypilot/SConscript
+++ b/sunnypilot/SConscript
@@ -0,0 +1 @@
+SConscript(['modeld/SConscript'])
--- a/sunnypilot/modeld/.gitignore
+++ b/sunnypilot/modeld/.gitignore
@@ -0,0 +1 @@
+*_pyx.cpp
--- a/sunnypilot/modeld/SConscript
+++ b/sunnypilot/modeld/SConscript
@@ -0,0 +1,58 @@
+import glob
+
+Import('env', 'envCython', 'arch', 'cereal', 'messaging', 'common', 'gpucommon', 'visionipc', 'transformations')
+lenv = env.Clone()
+lenvCython = envCython.Clone()
+
+libs = [cereal, messaging, visionipc, gpucommon, common, 'capnp', 'kj', 'pthread']
+frameworks = []
+
+common_src = [
+  "models/commonmodel.cc",
+  "transforms/loadyuv.cc",
+  "transforms/transform.cc",
+]
+
+thneed_src_common = [
+  "thneed/thneed_common.cc",
+  "thneed/serialize.cc",
+]
+
+thneed_src_qcom = thneed_src_common + ["thneed/thneed_qcom2.cc"]
+thneed_src_pc = thneed_src_common + ["thneed/thneed_pc.cc"]
+thneed_src = thneed_src_qcom if arch == "larch64" else thneed_src_pc
+
+# SNPE except on Mac and ARM Linux
+snpe_lib = []
+if arch != "Darwin" and arch != "aarch64":
+  common_src += ['runners/snpemodel.cc']
+  snpe_lib += ['SNPE']
+
+# OpenCL is a framework on Mac
+if arch == "Darwin":
+  frameworks += ['OpenCL']
+else:
+  libs += ['OpenCL']
+
+# Set path definitions
+for pathdef, fn in {'TRANSFORM': 'transforms/transform.cl', 'LOADYUV': 'transforms/loadyuv.cl'}.items():
+  for xenv in (lenv, lenvCython):
+    xenv['CXXFLAGS'].append(f'-D{pathdef}_PATH=\\"{File(fn).abspath}\\"')
+
+# Compile cython
+snpe_rpath_qcom = "/data/pythonpath/third_party/snpe/larch64"
+snpe_rpath_pc = f"{Dir('#').abspath}/third_party/snpe/x86_64-linux-clang"
+snpe_rpath = lenvCython['RPATH'] + [snpe_rpath_qcom if arch == "larch64" else snpe_rpath_pc]
+
+cython_libs = envCython["LIBS"] + libs
+snpemodel_lib = lenv.Library('snpemodel', ['runners/snpemodel.cc'])
+commonmodel_lib = lenv.Library('commonmodel', common_src)
+
+lenvCython.Program('runners/runmodel_pyx.so', 'runners/runmodel_pyx.pyx', LIBS=cython_libs, FRAMEWORKS=frameworks)
+lenvCython.Program('runners/snpemodel_pyx.so', 'runners/snpemodel_pyx.pyx', LIBS=[snpemodel_lib, snpe_lib, *cython_libs], FRAMEWORKS=frameworks, RPATH=snpe_rpath)
+lenvCython.Program('models/commonmodel_pyx.so', 'models/commonmodel_pyx.pyx', LIBS=[commonmodel_lib, *cython_libs], FRAMEWORKS=frameworks)
+
+if arch == 'larch64' or GetOption('pc_thneed'):
+  thneed_lib = env.SharedLibrary('thneed', thneed_src, LIBS=[gpucommon, common, 'OpenCL', 'dl'])
+  thneedmodel_lib = env.Library('thneedmodel', ['runners/thneedmodel.cc'])
+  lenvCython.Program('runners/thneedmodel_pyx.so', 'runners/thneedmodel_pyx.pyx', LIBS=envCython["LIBS"]+[thneedmodel_lib, thneed_lib, gpucommon, common, 'dl', 'OpenCL'])
--- a/sunnypilot/modeld/init.py
+++ b/sunnypilot/modeld/init.py
--- a/sunnypilot/modeld/constants.py
+++ b/sunnypilot/modeld/constants.py
@@ -0,0 +1,86 @@
+import numpy as np
+
+def index_function(idx, max_val=192, max_idx=32):
+  return (max_val) * ((idx/max_idx)**2)
+
+class ModelConstants:
+  # time and distance indices
+  IDX_N = 33
+  T_IDXS = [index_function(idx, max_val=10.0) for idx in range(IDX_N)]
+  X_IDXS = [index_function(idx, max_val=192.0) for idx in range(IDX_N)]
+  LEAD_T_IDXS = [0., 2., 4., 6., 8., 10.]
+  LEAD_T_OFFSETS = [0., 2., 4.]
+  META_T_IDXS = [2., 4., 6., 8., 10.]
+
+  # model inputs constants
+  MODEL_FREQ = 20
+  FEATURE_LEN = 512
+  FULL_HISTORY_BUFFER_LEN = 99
+  HISTORY_BUFFER_LEN = 24
+  DESIRE_LEN = 8
+  TRAFFIC_CONVENTION_LEN = 2
+  LAT_PLANNER_STATE_LEN = 4
+  LATERAL_CONTROL_PARAMS_LEN = 2
+  PREV_DESIRED_CURV_LEN = 1
+
+  # model outputs constants
+  FCW_THRESHOLDS_5MS2 = np.array([.05, .05, .15, .15, .15], dtype=np.float32)
+  FCW_THRESHOLDS_3MS2 = np.array([.7, .7], dtype=np.float32)
+  FCW_5MS2_PROBS_WIDTH = 5
+  FCW_3MS2_PROBS_WIDTH = 2
+
+  DISENGAGE_WIDTH = 5
+  POSE_WIDTH = 6
+  WIDE_FROM_DEVICE_WIDTH = 3
+  SIM_POSE_WIDTH = 6
+  LEAD_WIDTH = 4
+  LANE_LINES_WIDTH = 2
+  ROAD_EDGES_WIDTH = 2
+  PLAN_WIDTH = 15
+  DESIRE_PRED_WIDTH = 8
+  LAT_PLANNER_SOLUTION_WIDTH = 4
+  DESIRED_CURV_WIDTH = 1
+
+  NUM_LANE_LINES = 4
+  NUM_ROAD_EDGES = 2
+
+  LEAD_TRAJ_LEN = 6
+  DESIRE_PRED_LEN = 4
+
+  PLAN_MHP_N = 5
+  LEAD_MHP_N = 2
+  PLAN_MHP_SELECTION = 1
+  LEAD_MHP_SELECTION = 3
+
+  FCW_THRESHOLD_5MS2_HIGH = 0.15
+  FCW_THRESHOLD_5MS2_LOW = 0.05
+  FCW_THRESHOLD_3MS2 = 0.7
+
+  CONFIDENCE_BUFFER_LEN = 5
+  RYG_GREEN = 0.01165
+  RYG_YELLOW = 0.06157
+
+  POLY_PATH_DEGREE = 4
+
+# model outputs slices
+class Plan:
+  POSITION = slice(0, 3)
+  VELOCITY = slice(3, 6)
+  ACCELERATION = slice(6, 9)
+  T_FROM_CURRENT_EULER = slice(9, 12)
+  ORIENTATION_RATE = slice(12, 15)
+
+class Meta:
+  ENGAGED = slice(0, 1)
+  # next 2, 4, 6, 8, 10 seconds
+  GAS_DISENGAGE = slice(1, 31, 6)
+  BRAKE_DISENGAGE = slice(2, 31, 6)
+  STEER_OVERRIDE = slice(3, 31, 6)
+  HARD_BRAKE_3 = slice(4, 31, 6)
+  HARD_BRAKE_4 = slice(5, 31, 6)
+  HARD_BRAKE_5 = slice(6, 31, 6)
+  # next 0, 2, 4, 6, 8, 10 seconds
+  GAS_PRESS = slice(31, 55, 4)
+  BRAKE_PRESS = slice(32, 55, 4)
+  LEFT_BLINKER = slice(33, 55, 4)
+  RIGHT_BLINKER = slice(34, 55, 4)
--- a/sunnypilot/modeld/fill_model_msg.py
+++ b/sunnypilot/modeld/fill_model_msg.py
@@ -0,0 +1,237 @@
+import os
+import capnp
+import numpy as np
+from cereal import log
+from openpilot.sunnypilot.modeld.constants import ModelConstants, Plan, Meta
+from openpilot.selfdrive.controls.lib.drive_helpers import MIN_SPEED
+
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+
+ConfidenceClass = log.ModelDataV2.ConfidenceClass
+
+def curv_from_psis(psi_target, psi_rate, vego, delay):
+  vego = np.clip(vego, MIN_SPEED, np.inf)
+  curv_from_psi = psi_target / (vego * delay)  # epsilon to prevent divide-by-zero
+  return 2*curv_from_psi - psi_rate / vego
+
+def get_curvature_from_plan(plan, vego, delay):
+  psi_target = np.interp(delay, ModelConstants.T_IDXS, plan[:, Plan.T_FROM_CURRENT_EULER][:, 2])
+  psi_rate = plan[:, Plan.ORIENTATION_RATE][0, 2]
+  return curv_from_psis(psi_target, psi_rate, vego, delay)
+
+class PublishState:
+  def __init__(self):
+    self.disengage_buffer = np.zeros(ModelConstants.CONFIDENCE_BUFFER_LEN*ModelConstants.DISENGAGE_WIDTH, dtype=np.float32)
+    self.prev_brake_5ms2_probs = np.zeros(ModelConstants.FCW_5MS2_PROBS_WIDTH, dtype=np.float32)
+    self.prev_brake_3ms2_probs = np.zeros(ModelConstants.FCW_3MS2_PROBS_WIDTH, dtype=np.float32)
+
+def fill_xyzt(builder, t, x, y, z, x_std=None, y_std=None, z_std=None):
+  builder.t = t
+  builder.x = x.tolist()
+  builder.y = y.tolist()
+  builder.z = z.tolist()
+  if x_std is not None:
+    builder.xStd = x_std.tolist()
+  if y_std is not None:
+    builder.yStd = y_std.tolist()
+  if z_std is not None:
+    builder.zStd = z_std.tolist()
+
+def fill_xyvat(builder, t, x, y, v, a, x_std=None, y_std=None, v_std=None, a_std=None):
+  builder.t = t
+  builder.x = x.tolist()
+  builder.y = y.tolist()
+  builder.v = v.tolist()
+  builder.a = a.tolist()
+  if x_std is not None:
+    builder.xStd = x_std.tolist()
+  if y_std is not None:
+    builder.yStd = y_std.tolist()
+  if v_std is not None:
+    builder.vStd = v_std.tolist()
+  if a_std is not None:
+    builder.aStd = a_std.tolist()
+
+def fill_xyz_poly(builder, degree, x, y, z):
+  xyz = np.stack([x, y, z], axis=1)
+  coeffs = np.polynomial.polynomial.polyfit(ModelConstants.T_IDXS, xyz, deg=degree)
+  builder.xCoefficients = coeffs[:, 0].tolist()
+  builder.yCoefficients = coeffs[:, 1].tolist()
+  builder.zCoefficients = coeffs[:, 2].tolist()
+
+def fill_lane_line_meta(builder, lane_lines, lane_line_probs):
+  builder.leftY = lane_lines[1].y[0]
+  builder.leftProb = lane_line_probs[1]
+  builder.rightY = lane_lines[2].y[0]
+  builder.rightProb = lane_line_probs[2]
+
+def fill_model_msg(base_msg: capnp._DynamicStructBuilder, extended_msg: capnp._DynamicStructBuilder,
+                   net_output_data: dict[str, np.ndarray], v_ego: float, delay: float,
+                   publish_state: PublishState, vipc_frame_id: int, vipc_frame_id_extra: int,
+                   frame_id: int, frame_drop: float, timestamp_eof: int, model_execution_time: float,
+                   valid: bool) -> None:
+  frame_age = frame_id - vipc_frame_id if frame_id > vipc_frame_id else 0
+  frame_drop_perc = frame_drop * 100
+  extended_msg.valid = valid
+  base_msg.valid = valid
+
+  desired_curv = float(get_curvature_from_plan(net_output_data['plan'][0], v_ego, delay))
+
+  driving_model_data = base_msg.drivingModelData
+
+  driving_model_data.frameId = vipc_frame_id
+  driving_model_data.frameIdExtra = vipc_frame_id_extra
+  driving_model_data.frameDropPerc = frame_drop_perc
+  driving_model_data.modelExecutionTime = model_execution_time
+
+  action = driving_model_data.action
+  action.desiredCurvature = desired_curv
+
+  modelV2 = extended_msg.modelV2
+  modelV2.frameId = vipc_frame_id
+  modelV2.frameIdExtra = vipc_frame_id_extra
+  modelV2.frameAge = frame_age
+  modelV2.frameDropPerc = frame_drop_perc
+  modelV2.timestampEof = timestamp_eof
+  modelV2.modelExecutionTime = model_execution_time
+
+  # plan
+  position = modelV2.position
+  fill_xyzt(position, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.POSITION].T, *net_output_data['plan_stds'][0,:,Plan.POSITION].T)
+  velocity = modelV2.velocity
+  fill_xyzt(velocity, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.VELOCITY].T)
+  acceleration = modelV2.acceleration
+  fill_xyzt(acceleration, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ACCELERATION].T)
+  orientation = modelV2.orientation
+  fill_xyzt(orientation, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.T_FROM_CURRENT_EULER].T)
+  orientation_rate = modelV2.orientationRate
+  fill_xyzt(orientation_rate, ModelConstants.T_IDXS, *net_output_data['plan'][0,:,Plan.ORIENTATION_RATE].T)
+
+  # temporal pose
+  temporal_pose = modelV2.temporalPose
+  temporal_pose.trans = net_output_data['plan'][0,0,Plan.VELOCITY].tolist()
+  temporal_pose.transStd = net_output_data['plan_stds'][0,0,Plan.VELOCITY].tolist()
+  temporal_pose.rot = net_output_data['plan'][0,0,Plan.ORIENTATION_RATE].tolist()
+  temporal_pose.rotStd = net_output_data['plan_stds'][0,0,Plan.ORIENTATION_RATE].tolist()
+
+  # poly path
+  poly_path = driving_model_data.path
+  fill_xyz_poly(poly_path, ModelConstants.POLY_PATH_DEGREE, *net_output_data['plan'][0,:,Plan.POSITION].T)
+
+  # lateral planning
+  action = modelV2.action
+  action.desiredCurvature = desired_curv
+
+  # times at X_IDXS according to model plan
+  PLAN_T_IDXS = [np.nan] * ModelConstants.IDX_N
+  PLAN_T_IDXS[0] = 0.0
+  plan_x = net_output_data['plan'][0,:,Plan.POSITION][:,0].tolist()
+  for xidx in range(1, ModelConstants.IDX_N):
+    tidx = 0
+    # increment tidx until we find an element that's further away than the current xidx
+    while tidx < ModelConstants.IDX_N - 1 and plan_x[tidx+1] < ModelConstants.X_IDXS[xidx]:
+      tidx += 1
+    if tidx == ModelConstants.IDX_N - 1:
+      # if the Plan doesn't extend far enough, set plan_t to the max value (10s), then break
+      PLAN_T_IDXS[xidx] = ModelConstants.T_IDXS[ModelConstants.IDX_N - 1]
+      break
+    # interpolate to find `t` for the current xidx
+    current_x_val = plan_x[tidx]
+    next_x_val = plan_x[tidx+1]
+    p = (ModelConstants.X_IDXS[xidx] - current_x_val) / (next_x_val - current_x_val) if abs(next_x_val - current_x_val) > 1e-9 else float('nan')
+    PLAN_T_IDXS[xidx] = p * ModelConstants.T_IDXS[tidx+1] + (1 - p) * ModelConstants.T_IDXS[tidx]
+
+  # lane lines
+  modelV2.init('laneLines', 4)
+  for i in range(4):
+    lane_line = modelV2.laneLines[i]
+    fill_xyzt(lane_line, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), net_output_data['lane_lines'][0,i,:,0], net_output_data['lane_lines'][0,i,:,1])
+  modelV2.laneLineStds = net_output_data['lane_lines_stds'][0,:,0,0].tolist()
+  modelV2.laneLineProbs = net_output_data['lane_lines_prob'][0,1::2].tolist()
+
+  lane_line_meta = driving_model_data.laneLineMeta
+  fill_lane_line_meta(lane_line_meta, modelV2.laneLines, modelV2.laneLineProbs)
+
+  # road edges
+  modelV2.init('roadEdges', 2)
+  for i in range(2):
+    road_edge = modelV2.roadEdges[i]
+    fill_xyzt(road_edge, PLAN_T_IDXS, np.array(ModelConstants.X_IDXS), net_output_data['road_edges'][0,i,:,0], net_output_data['road_edges'][0,i,:,1])
+  modelV2.roadEdgeStds = net_output_data['road_edges_stds'][0,:,0,0].tolist()
+
+  # leads
+  modelV2.init('leadsV3', 3)
+  for i in range(3):
+    lead = modelV2.leadsV3[i]
+    fill_xyvat(lead, ModelConstants.LEAD_T_IDXS, *net_output_data['lead'][0,i].T, *net_output_data['lead_stds'][0,i].T)
+    lead.prob = net_output_data['lead_prob'][0,i].tolist()
+    lead.probTime = ModelConstants.LEAD_T_OFFSETS[i]
+
+  # meta
+  meta = modelV2.meta
+  meta.desireState = net_output_data['desire_state'][0].reshape(-1).tolist()
+  meta.desirePrediction = net_output_data['desire_pred'][0].reshape(-1).tolist()
+  meta.engagedProb = net_output_data['meta'][0,Meta.ENGAGED].item()
+  meta.init('disengagePredictions')
+  disengage_predictions = meta.disengagePredictions
+  disengage_predictions.t = ModelConstants.META_T_IDXS
+  disengage_predictions.brakeDisengageProbs = net_output_data['meta'][0,Meta.BRAKE_DISENGAGE].tolist()
+  disengage_predictions.gasDisengageProbs = net_output_data['meta'][0,Meta.GAS_DISENGAGE].tolist()
+  disengage_predictions.steerOverrideProbs = net_output_data['meta'][0,Meta.STEER_OVERRIDE].tolist()
+  disengage_predictions.brake3MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_3].tolist()
+  disengage_predictions.brake4MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_4].tolist()
+  disengage_predictions.brake5MetersPerSecondSquaredProbs = net_output_data['meta'][0,Meta.HARD_BRAKE_5].tolist()
+  disengage_predictions.gasPressProbs = net_output_data['meta'][0,Meta.GAS_PRESS].tolist()
+  disengage_predictions.brakePressProbs = net_output_data['meta'][0,Meta.BRAKE_PRESS].tolist()
+
+  publish_state.prev_brake_5ms2_probs[:-1] = publish_state.prev_brake_5ms2_probs[1:]
+  publish_state.prev_brake_5ms2_probs[-1] = net_output_data['meta'][0,Meta.HARD_BRAKE_5][0]
+  publish_state.prev_brake_3ms2_probs[:-1] = publish_state.prev_brake_3ms2_probs[1:]
+  publish_state.prev_brake_3ms2_probs[-1] = net_output_data['meta'][0,Meta.HARD_BRAKE_3][0]
+  hard_brake_predicted = (publish_state.prev_brake_5ms2_probs > ModelConstants.FCW_THRESHOLDS_5MS2).all() and \
+    (publish_state.prev_brake_3ms2_probs > ModelConstants.FCW_THRESHOLDS_3MS2).all()
+  meta.hardBrakePredicted = hard_brake_predicted.item()
+
+  # confidence
+  if vipc_frame_id % (2*ModelConstants.MODEL_FREQ) == 0:
+    # any disengage prob
+    brake_disengage_probs = net_output_data['meta'][0,Meta.BRAKE_DISENGAGE]
+    gas_disengage_probs = net_output_data['meta'][0,Meta.GAS_DISENGAGE]
+    steer_override_probs = net_output_data['meta'][0,Meta.STEER_OVERRIDE]
+    any_disengage_probs = 1-((1-brake_disengage_probs)*(1-gas_disengage_probs)*(1-steer_override_probs))
+    # independent disengage prob for each 2s slice
+    ind_disengage_probs = np.r_[any_disengage_probs[0], np.diff(any_disengage_probs) / (1 - any_disengage_probs[:-1])]
+    # rolling buf for 2, 4, 6, 8, 10s
+    publish_state.disengage_buffer[:-ModelConstants.DISENGAGE_WIDTH] = publish_state.disengage_buffer[ModelConstants.DISENGAGE_WIDTH:]
+    publish_state.disengage_buffer[-ModelConstants.DISENGAGE_WIDTH:] = ind_disengage_probs
+
+  score = 0.
+  for i in range(ModelConstants.DISENGAGE_WIDTH):
+    score += publish_state.disengage_buffer[i*ModelConstants.DISENGAGE_WIDTH+ModelConstants.DISENGAGE_WIDTH-1-i].item() / ModelConstants.DISENGAGE_WIDTH
+  if score < ModelConstants.RYG_GREEN:
+    modelV2.confidence = ConfidenceClass.green
+  elif score < ModelConstants.RYG_YELLOW:
+    modelV2.confidence = ConfidenceClass.yellow
+  else:
+    modelV2.confidence = ConfidenceClass.red
+
+  # raw prediction if enabled
+  if SEND_RAW_PRED:
+    modelV2.rawPredictions = net_output_data['raw_pred'].tobytes()
+
+def fill_pose_msg(msg: capnp._DynamicStructBuilder, net_output_data: dict[str, np.ndarray],
+                  vipc_frame_id: int, vipc_dropped_frames: int, timestamp_eof: int, live_calib_seen: bool) -> None:
+  msg.valid = live_calib_seen & (vipc_dropped_frames < 1)
+  cameraOdometry = msg.cameraOdometry
+
+  cameraOdometry.frameId = vipc_frame_id
+  cameraOdometry.timestampEof = timestamp_eof
+
+  cameraOdometry.trans = net_output_data['pose'][0,:3].tolist()
+  cameraOdometry.rot = net_output_data['pose'][0,3:].tolist()
+  cameraOdometry.wideFromDeviceEuler = net_output_data['wide_from_device_euler'][0,:].tolist()
+  cameraOdometry.roadTransformTrans = net_output_data['road_transform'][0,:3].tolist()
+  cameraOdometry.transStd = net_output_data['pose_stds'][0,:3].tolist()
+  cameraOdometry.rotStd = net_output_data['pose_stds'][0,3:].tolist()
+  cameraOdometry.wideFromDeviceEulerStd = net_output_data['wide_from_device_euler_stds'][0,:].tolist()
+  cameraOdometry.roadTransformTransStd = net_output_data['road_transform_stds'][0,:3].tolist()
--- a/sunnypilot/modeld/get_model_metadata.py
+++ b/sunnypilot/modeld/get_model_metadata.py
@@ -0,0 +1,28 @@
+#!/usr/bin/env python3
+import sys
+import pathlib
+import onnx
+import codecs
+import pickle
+
+def get_name_and_shape(value_info:onnx.ValueInfoProto) -> tuple[str, tuple[int,...]]:
+  shape = tuple([int(dim.dim_value) for dim in value_info.type.tensor_type.shape.dim])
+  name = value_info.name
+  return name, shape
+
+if __name__ == "__main__":
+  model_path = pathlib.Path(sys.argv[1])
+  model = onnx.load(str(model_path))
+  i = [x.key for x in model.metadata_props].index('output_slices')
+  output_slices = model.metadata_props[i].value
+
+  metadata = {}
+  metadata['output_slices'] = pickle.loads(codecs.decode(output_slices.encode(), "base64"))
+  metadata['input_shapes'] = dict([get_name_and_shape(x) for x in model.graph.input])
+  metadata['output_shapes'] = dict([get_name_and_shape(x) for x in model.graph.output])
+
+  metadata_path = model_path.parent / (model_path.stem + '_metadata.pkl')
+  with open(metadata_path, 'wb') as f:
+    pickle.dump(metadata, f)
+
+  print(f'saved metadata to {metadata_path}')
--- a/sunnypilot/modeld/modeld
+++ b/sunnypilot/modeld/modeld
@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" >/dev/null && pwd)"
+cd "$DIR/../../"
+
+if [ -f "$DIR/libthneed.so" ]; then
+  export LD_PRELOAD="$DIR/libthneed.so"
+fi
+
+exec "$DIR/modeld.py" "$@"
--- a/sunnypilot/modeld/modeld.py
+++ b/sunnypilot/modeld/modeld.py
@@ -0,0 +1,299 @@
+#!/usr/bin/env python3
+import os
+import time
+import pickle
+import numpy as np
+import cereal.messaging as messaging
+from cereal import car, log
+from pathlib import Path
+from setproctitle import setproctitle
+from cereal.messaging import PubMaster, SubMaster
+from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
+from opendbc.car.car_helpers import get_demo_car_params
+from openpilot.common.swaglog import cloudlog
+from openpilot.common.params import Params
+from openpilot.common.filter_simple import FirstOrderFilter
+from openpilot.common.realtime import config_realtime_process
+from openpilot.common.transformations.camera import DEVICE_CAMERAS
+from openpilot.common.transformations.model import get_warp_matrix
+from openpilot.system import sentry
+from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
+from openpilot.sunnypilot.modeld.runners import ModelRunner, Runtime
+from openpilot.sunnypilot.modeld.parse_model_outputs import Parser
+from openpilot.sunnypilot.modeld.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState
+from openpilot.sunnypilot.modeld.constants import ModelConstants
+from openpilot.sunnypilot.modeld.models.commonmodel_pyx import DrivingModelFrame, CLContext
+
+PROCESS_NAME = "sunnypilot.modeld.modeld"
+SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
+
+MODEL_PATHS = {
+  ModelRunner.THNEED: Path(__file__).parent / 'models/supercombo.thneed',
+  ModelRunner.ONNX: Path(__file__).parent / 'models/supercombo.onnx'}
+
+METADATA_PATH = Path(__file__).parent / 'models/supercombo_metadata.pkl'
+
+
+class FrameMeta:
+  frame_id: int = 0
+  timestamp_sof: int = 0
+  timestamp_eof: int = 0
+
+  def __init__(self, vipc=None):
+    if vipc is not None:
+      self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
+
+class ModelState:
+  frame: DrivingModelFrame
+  wide_frame: DrivingModelFrame
+  inputs: dict[str, np.ndarray]
+  output: np.ndarray
+  prev_desire: np.ndarray  # for tracking the rising edge of the pulse
+  model: ModelRunner
+
+  def __init__(self, context: CLContext):
+    self.frame = DrivingModelFrame(context)
+    self.wide_frame = DrivingModelFrame(context)
+    self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
+    self.full_features_20Hz = np.zeros((ModelConstants.FULL_HISTORY_BUFFER_LEN, ModelConstants.FEATURE_LEN), dtype=np.float32)
+    self.desire_20Hz =  np.zeros((ModelConstants.FULL_HISTORY_BUFFER_LEN + 1, ModelConstants.DESIRE_LEN), dtype=np.float32)
+
+    # img buffers are managed in openCL transform code
+    self.inputs = {
+      'desire': np.zeros(ModelConstants.DESIRE_LEN * (ModelConstants.HISTORY_BUFFER_LEN+1), dtype=np.float32),
+      'traffic_convention': np.zeros(ModelConstants.TRAFFIC_CONVENTION_LEN, dtype=np.float32),
+      'features_buffer': np.zeros(ModelConstants.HISTORY_BUFFER_LEN * ModelConstants.FEATURE_LEN, dtype=np.float32),
+    }
+
+    with open(METADATA_PATH, 'rb') as f:
+      model_metadata = pickle.load(f)
+
+    self.output_slices = model_metadata['output_slices']
+    net_output_size = model_metadata['output_shapes']['outputs'][1]
+    self.output = np.zeros(net_output_size, dtype=np.float32)
+    self.parser = Parser()
+
+    self.model = ModelRunner(MODEL_PATHS, self.output, Runtime.GPU, False, context)
+    self.model.addInput("input_imgs", None)
+    self.model.addInput("big_input_imgs", None)
+    for k,v in self.inputs.items():
+      self.model.addInput(k, v)
+
+  def slice_outputs(self, model_outputs: np.ndarray) -> dict[str, np.ndarray]:
+    parsed_model_outputs = {k: model_outputs[np.newaxis, v] for k,v in self.output_slices.items()}
+    if SEND_RAW_PRED:
+      parsed_model_outputs['raw_pred'] = model_outputs.copy()
+    return parsed_model_outputs
+
+  def run(self, buf: VisionBuf, wbuf: VisionBuf, transform: np.ndarray, transform_wide: np.ndarray,
+                inputs: dict[str, np.ndarray], prepare_only: bool) -> dict[str, np.ndarray] | None:
+    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
+    inputs['desire'][0] = 0
+    new_desire = np.where(inputs['desire'] - self.prev_desire > .99, inputs['desire'], 0)
+    self.prev_desire[:] = inputs['desire']
+
+    self.desire_20Hz[:-1] = self.desire_20Hz[1:]
+    self.desire_20Hz[-1] = new_desire
+    self.inputs['desire'][:] = self.desire_20Hz.reshape((25,4,-1)).max(axis=1).flatten()
+
+    self.inputs['traffic_convention'][:] = inputs['traffic_convention']
+
+    self.model.setInputBuffer("input_imgs", self.frame.prepare(buf, transform.flatten(), self.model.getCLBuffer("input_imgs")))
+    self.model.setInputBuffer("big_input_imgs", self.wide_frame.prepare(wbuf, transform_wide.flatten(), self.model.getCLBuffer("big_input_imgs")))
+
+    if prepare_only:
+      return None
+
+    self.model.execute()
+    outputs = self.parser.parse_outputs(self.slice_outputs(self.output))
+
+    self.full_features_20Hz[:-1] = self.full_features_20Hz[1:]
+    self.full_features_20Hz[-1] = outputs['hidden_state'][0, :]
+
+    idxs = np.arange(-4,-100,-4)[::-1]
+    self.inputs['features_buffer'][:] = self.full_features_20Hz[idxs].flatten()
+    return outputs
+
+
+def main(demo=False):
+  cloudlog.warning("modeld init")
+
+  sentry.set_tag("daemon", PROCESS_NAME)
+  cloudlog.bind(daemon=PROCESS_NAME)
+  setproctitle(PROCESS_NAME)
+  config_realtime_process(7, 54)
+
+  cloudlog.warning("setting up CL context")
+  cl_context = CLContext()
+  cloudlog.warning("CL context ready; loading model")
+  model = ModelState(cl_context)
+  cloudlog.warning("models loaded, modeld starting")
+
+  # visionipc clients
+  while True:
+    available_streams = VisionIpcClient.available_streams("camerad", block=False)
+    if available_streams:
+      use_extra_client = VisionStreamType.VISION_STREAM_WIDE_ROAD in available_streams and VisionStreamType.VISION_STREAM_ROAD in available_streams
+      main_wide_camera = VisionStreamType.VISION_STREAM_ROAD not in available_streams
+      break
+    time.sleep(.1)
+
+  vipc_client_main_stream = VisionStreamType.VISION_STREAM_WIDE_ROAD if main_wide_camera else VisionStreamType.VISION_STREAM_ROAD
+  vipc_client_main = VisionIpcClient("camerad", vipc_client_main_stream, True, cl_context)
+  vipc_client_extra = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_WIDE_ROAD, False, cl_context)
+  cloudlog.warning(f"vision stream set up, main_wide_camera: {main_wide_camera}, use_extra_client: {use_extra_client}")
+
+  while not vipc_client_main.connect(False):
+    time.sleep(0.1)
+  while use_extra_client and not vipc_client_extra.connect(False):
+    time.sleep(0.1)
+
+  cloudlog.warning(f"connected main cam with buffer size: {vipc_client_main.buffer_len} ({vipc_client_main.width} x {vipc_client_main.height})")
+  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})")
+
+  # messaging
+  pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry"])
+  sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl"])
+
+  publish_state = PublishState()
+  params = Params()
+
+  # setup filter to track dropped frames
+  frame_dropped_filter = FirstOrderFilter(0., 10., 1. / ModelConstants.MODEL_FREQ)
+  frame_id = 0
+  last_vipc_frame_id = 0
+  run_count = 0
+
+  model_transform_main = np.zeros((3, 3), dtype=np.float32)
+  model_transform_extra = np.zeros((3, 3), dtype=np.float32)
+  live_calib_seen = False
+  buf_main, buf_extra = None, None
+  meta_main = FrameMeta()
+  meta_extra = FrameMeta()
+
+
+  if demo:
+    CP = get_demo_car_params()
+  else:
+    CP = messaging.log_from_bytes(params.get("CarParams", block=True), car.CarParams)
+  cloudlog.info("modeld got CarParams: %s", CP.carName)
+
+  # TODO this needs more thought, use .2s extra for now to estimate other delays
+  steer_delay = CP.steerActuatorDelay + .2
+
+  DH = DesireHelper()
+
+  while True:
+    # Keep receiving frames until we are at least 1 frame ahead of previous extra frame
+    while meta_main.timestamp_sof < meta_extra.timestamp_sof + 25000000:
+      buf_main = vipc_client_main.recv()
+      meta_main = FrameMeta(vipc_client_main)
+      if buf_main is None:
+        break
+
+    if buf_main is None:
+      cloudlog.debug("vipc_client_main no frame")
+      continue
+
+    if use_extra_client:
+      # Keep receiving extra frames until frame id matches main camera
+      while True:
+        buf_extra = vipc_client_extra.recv()
+        meta_extra = FrameMeta(vipc_client_extra)
+        if buf_extra is None or meta_main.timestamp_sof < meta_extra.timestamp_sof + 25000000:
+          break
+
+      if buf_extra is None:
+        cloudlog.debug("vipc_client_extra no frame")
+        continue
+
+      if abs(meta_main.timestamp_sof - meta_extra.timestamp_sof) > 10000000:
+        cloudlog.error(f"frames out of sync! main: {meta_main.frame_id} ({meta_main.timestamp_sof / 1e9:.5f}),\
+                         extra: {meta_extra.frame_id} ({meta_extra.timestamp_sof / 1e9:.5f})")
+
+    else:
+      # Use single camera
+      buf_extra = buf_main
+      meta_extra = meta_main
+
+    sm.update(0)
+    desire = DH.desire
+    is_rhd = sm["driverMonitoringState"].isRHD
+    frame_id = sm["roadCameraState"].frameId
+    v_ego = max(sm["carState"].vEgo, 0.)
+    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))]
+      model_transform_main = get_warp_matrix(device_from_calib_euler, dc.ecam.intrinsics if main_wide_camera else dc.fcam.intrinsics, False).astype(np.float32)
+      model_transform_extra = get_warp_matrix(device_from_calib_euler, dc.ecam.intrinsics, True).astype(np.float32)
+      live_calib_seen = True
+
+    traffic_convention = np.zeros(2)
+    traffic_convention[int(is_rhd)] = 1
+
+    vec_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
+    if desire >= 0 and desire < ModelConstants.DESIRE_LEN:
+      vec_desire[desire] = 1
+
+    # tracked dropped frames
+    vipc_dropped_frames = max(0, meta_main.frame_id - last_vipc_frame_id - 1)
+    frames_dropped = frame_dropped_filter.update(min(vipc_dropped_frames, 10))
+    if run_count < 10: # let frame drops warm up
+      frame_dropped_filter.x = 0.
+      frames_dropped = 0.
+    run_count = run_count + 1
+
+    frame_drop_ratio = frames_dropped / (1 + frames_dropped)
+    prepare_only = vipc_dropped_frames > 0
+    if prepare_only:
+      cloudlog.error(f"skipping model eval. Dropped {vipc_dropped_frames} frames")
+
+    inputs:dict[str, np.ndarray] = {
+      'desire': vec_desire,
+      'traffic_convention': traffic_convention,
+      }
+
+    mt1 = time.perf_counter()
+    model_output = model.run(buf_main, buf_extra, model_transform_main, model_transform_extra, inputs, prepare_only)
+    mt2 = time.perf_counter()
+    model_execution_time = mt2 - mt1
+
+    if model_output is not None:
+      modelv2_send = messaging.new_message('modelV2')
+      drivingdata_send = messaging.new_message('drivingModelData')
+      posenet_send = messaging.new_message('cameraOdometry')
+      fill_model_msg(drivingdata_send, modelv2_send, model_output, v_ego, steer_delay,
+                     publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id,
+                     frame_drop_ratio, meta_main.timestamp_eof, model_execution_time, live_calib_seen)
+
+      desire_state = modelv2_send.modelV2.meta.desireState
+      l_lane_change_prob = desire_state[log.Desire.laneChangeLeft]
+      r_lane_change_prob = desire_state[log.Desire.laneChangeRight]
+      lane_change_prob = l_lane_change_prob + r_lane_change_prob
+      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
+      drivingdata_send.drivingModelData.meta.laneChangeState = DH.lane_change_state
+      drivingdata_send.drivingModelData.meta.laneChangeDirection = DH.lane_change_direction
+
+      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)
+
+    last_vipc_frame_id = meta_main.frame_id
+
+
+if __name__ == "__main__":
+  try:
+    import argparse
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--demo', action='store_true', help='A boolean for demo mode.')
+    args = parser.parse_args()
+    main(demo=args.demo)
+  except KeyboardInterrupt:
+    cloudlog.warning(f"child {PROCESS_NAME} got SIGINT")
+  except Exception:
+    sentry.capture_exception()
+    raise
--- a/sunnypilot/modeld/models/README.md
+++ b/sunnypilot/modeld/models/README.md
@@ -0,0 +1,62 @@
+## Neural networks in openpilot
+To view the architecture of the ONNX networks, you can use [netron](https://netron.app/)
+
+## Supercombo
+### Supercombo input format (Full size: 799906 x float32)
+* **image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+* **wide image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+* **desire**
+  * one-hot encoded buffer to command model to execute certain actions, bit needs to be sent for the past 5 seconds (at 20FPS) : 100 * 8
+* **traffic convention**
+  * one-hot encoded vector to tell model whether traffic is right-hand or left-hand traffic : 2
+* **feature buffer**
+  * A buffer of intermediate features that gets appended to the current feature to form a 5 seconds temporal context (at 20FPS) : 99 * 512
+
+
+### Supercombo output format (Full size: XXX x float32)
+Read [here](https://github.com/commaai/openpilot/blob/90af436a121164a51da9fa48d093c29f738adf6a/selfdrive/modeld/models/driving.h#L236) for more.
+
+
+## Driver Monitoring Model
+* .onnx model can be run with onnx runtimes
+* .dlc file is a pre-quantized model and only runs on qualcomm DSPs
+
+### input format
+* single image W = 1440 H = 960 luminance channel (Y) from the planar YUV420 format:
+  * full input size is 1440 * 960 = 1382400
+  * normalized ranging from 0.0 to 1.0 in float32 (onnx runner) or ranging from 0 to 255 in uint8 (snpe runner)
+* camera calibration angles (roll, pitch, yaw) from liveCalibration: 3 x float32 inputs
+
+### output format
+* 84 x float32 outputs = 2 + 41 * 2 ([parsing example](https://github.com/commaai/openpilot/blob/22ce4e17ba0d3bfcf37f8255a4dd1dc683fe0c38/selfdrive/modeld/models/dmonitoring.cc#L33))
+  * for each person in the front seats (2 * 41)
+    * face pose: 12 = 6 + 6
+      * face orientation [pitch, yaw, roll] in camera frame: 3
+      * face position [dx, dy] relative to image center: 2
+      * normalized face size: 1
+      * standard deviations for above outputs: 6
+    * face visible probability: 1
+    * eyes: 20 = (8 + 1) + (8 + 1) + 1 + 1
+      * eye position and size, and their standard deviations: 8
+      * eye visible probability: 1
+      * eye closed probability: 1
+    * wearing sunglasses probability: 1
+    * face occluded probability: 1
+    * touching wheel probability: 1
+    * paying attention probability: 1
+    * (deprecated) distracted probabilities: 2
+    * using phone probability: 1
+    * distracted probability: 1
+  * common outputs 2
+    * poor camera vision probability: 1
+    * left hand drive probability: 1
--- a/sunnypilot/modeld/models/commonmodel.cc
+++ b/sunnypilot/modeld/models/commonmodel.cc
@@ -0,0 +1,69 @@
+#include "sunnypilot/modeld/models/commonmodel.h"
+
+#include <cmath>
+#include <cstring>
+
+#include "common/clutil.h"
+
+DrivingModelFrame::DrivingModelFrame(cl_device_id device_id, cl_context context) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  //input_frames_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+  img_buffer_20hz_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, 5*frame_size_bytes, NULL, &err));
+  region.origin = 4 * frame_size_bytes;
+  region.size = frame_size_bytes;
+  last_img_cl = CL_CHECK_ERR(clCreateSubBuffer(img_buffer_20hz_cl, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err));
+
+  loadyuv_init(&loadyuv, context, device_id, MODEL_WIDTH, MODEL_HEIGHT);
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+
+uint8_t* DrivingModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection, cl_mem* output) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+
+  for (int i = 0; i < 4; i++) {
+    CL_CHECK(clEnqueueCopyBuffer(q, img_buffer_20hz_cl, img_buffer_20hz_cl, (i+1)*frame_size_bytes, i*frame_size_bytes, frame_size_bytes, 0, nullptr, nullptr));
+  }
+  loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, last_img_cl);
+
+  if (output == NULL) {
+    CL_CHECK(clEnqueueReadBuffer(q, img_buffer_20hz_cl, CL_TRUE, 0, frame_size_bytes, &input_frames[0], 0, nullptr, nullptr));
+    CL_CHECK(clEnqueueReadBuffer(q, last_img_cl, CL_TRUE, 0, frame_size_bytes, &input_frames[MODEL_FRAME_SIZE], 0, nullptr, nullptr));
+    clFinish(q);
+    return &input_frames[0];
+  } else {
+    copy_queue(&loadyuv, q, img_buffer_20hz_cl, *output, 0, 0, frame_size_bytes);
+    copy_queue(&loadyuv, q, last_img_cl, *output, 0, frame_size_bytes, frame_size_bytes);
+
+    // NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready.
+    clFinish(q);
+    return NULL;
+  }
+}
+
+DrivingModelFrame::~DrivingModelFrame() {
+  deinit_transform();
+  loadyuv_destroy(&loadyuv);
+  CL_CHECK(clReleaseMemObject(img_buffer_20hz_cl));
+  CL_CHECK(clReleaseMemObject(last_img_cl));
+  CL_CHECK(clReleaseCommandQueue(q));
+}
+
+
+MonitoringModelFrame::MonitoringModelFrame(cl_device_id device_id, cl_context context) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  //input_frame_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+uint8_t* MonitoringModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection, cl_mem* output) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+  CL_CHECK(clEnqueueReadBuffer(q, y_cl, CL_TRUE, 0, MODEL_FRAME_SIZE * sizeof(uint8_t), input_frames.get(), 0, nullptr, nullptr));
+  clFinish(q);
+  //return &y_cl;
+  return input_frames.get();
+}
+
+MonitoringModelFrame::~MonitoringModelFrame() {
+  deinit_transform();
+  CL_CHECK(clReleaseCommandQueue(q));
+}
--- a/sunnypilot/modeld/models/commonmodel.h
+++ b/sunnypilot/modeld/models/commonmodel.h
@@ -0,0 +1,98 @@
+#pragma once
+
+#include <cfloat>
+#include <cstdlib>
+#include <cassert>
+
+#include <memory>
+
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#ifdef __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+#include "common/mat.h"
+#include "selfdrive/modeld/transforms/loadyuv.h"
+#include "selfdrive/modeld/transforms/transform.h"
+
+class ModelFrame {
+public:
+  ModelFrame(cl_device_id device_id, cl_context context) {
+    q = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &err));
+  }
+  virtual ~ModelFrame() {}
+  virtual uint8_t* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection, cl_mem* output) { return NULL; }
+  /*
+  uint8_t* buffer_from_cl(cl_mem *in_frames, int buffer_size) {
+    CL_CHECK(clEnqueueReadBuffer(q, *in_frames, CL_TRUE, 0, buffer_size, input_frames.get(), 0, nullptr, nullptr));
+    clFinish(q);
+    return &input_frames[0];
+  }
+  */
+
+  int MODEL_WIDTH;
+  int MODEL_HEIGHT;
+  int MODEL_FRAME_SIZE;
+  int buf_size;
+
+protected:
+  cl_mem y_cl, u_cl, v_cl;
+  Transform transform;
+  cl_command_queue q;
+  std::unique_ptr<uint8_t[]> input_frames;
+
+  void init_transform(cl_device_id device_id, cl_context context, int model_width, int model_height) {
+    y_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, model_width * model_height, NULL, &err));
+    u_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    v_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    transform_init(&transform, context, device_id);
+  }
+
+  void deinit_transform() {
+    transform_destroy(&transform);
+    CL_CHECK(clReleaseMemObject(v_cl));
+    CL_CHECK(clReleaseMemObject(u_cl));
+    CL_CHECK(clReleaseMemObject(y_cl));
+  }
+
+  void run_transform(cl_mem yuv_cl, int model_width, int model_height, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+    transform_queue(&transform, q,
+        yuv_cl, frame_width, frame_height, frame_stride, frame_uv_offset,
+        y_cl, u_cl, v_cl, model_width, model_height, projection);
+  }
+};
+
+class DrivingModelFrame : public ModelFrame {
+public:
+  DrivingModelFrame(cl_device_id device_id, cl_context context);
+  ~DrivingModelFrame();
+  uint8_t* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection, cl_mem* output);
+
+  const int MODEL_WIDTH = 512;
+  const int MODEL_HEIGHT = 256;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT * 3 / 2;
+  const int buf_size = MODEL_FRAME_SIZE * 2;
+  const size_t frame_size_bytes = MODEL_FRAME_SIZE * sizeof(uint8_t);
+
+private:
+  LoadYUVState loadyuv;
+  cl_mem img_buffer_20hz_cl, last_img_cl;//, input_frames_cl;
+  cl_buffer_region region;
+};
+
+class MonitoringModelFrame : public ModelFrame {
+public:
+  MonitoringModelFrame(cl_device_id device_id, cl_context context);
+  ~MonitoringModelFrame();
+  uint8_t* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection, cl_mem* output);
+
+  const int MODEL_WIDTH = 1440;
+  const int MODEL_HEIGHT = 960;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT;
+  const int buf_size = MODEL_FRAME_SIZE;
+
+private:
+  // cl_mem input_frame_cl;
+};
--- a/sunnypilot/modeld/models/commonmodel.pxd
+++ b/sunnypilot/modeld/models/commonmodel.pxd
@@ -0,0 +1,26 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_device_id, cl_context, cl_mem
+
+cdef extern from "common/mat.h":
+  cdef struct mat3:
+    float v[9]
+
+cdef extern from "common/clutil.h":
+  cdef unsigned long CL_DEVICE_TYPE_DEFAULT
+  cl_device_id cl_get_device_id(unsigned long)
+  cl_context cl_create_context(cl_device_id)
+
+cdef extern from "sunnypilot/modeld/models/commonmodel.h":
+  cppclass ModelFrame:
+    int buf_size
+    # unsigned char * buffer_from_cl(cl_mem*, int);
+    unsigned char * prepare(cl_mem, int, int, int, int, mat3, cl_mem*)
+
+  cppclass DrivingModelFrame:
+    int buf_size
+    DrivingModelFrame(cl_device_id, cl_context)
+
+  cppclass MonitoringModelFrame:
+    int buf_size
+    MonitoringModelFrame(cl_device_id, cl_context)
--- a/sunnypilot/modeld/models/commonmodel_pyx.pxd
+++ b/sunnypilot/modeld/models/commonmodel_pyx.pxd
@@ -0,0 +1,13 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport CLContext as BaseCLContext
+
+cdef class CLContext(BaseCLContext):
+  pass
+
+cdef class CLMem:
+  cdef cl_mem * mem
+
+  @staticmethod
+  cdef create(void*)
--- a/sunnypilot/modeld/models/commonmodel_pyx.pyx
+++ b/sunnypilot/modeld/models/commonmodel_pyx.pyx
@@ -0,0 +1,76 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+import numpy as np
+cimport numpy as cnp
+from libc.string cimport memcpy
+from libc.stdint cimport uintptr_t
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport VisionBuf, CLContext as BaseCLContext
+from sunnypilot.modeld.models.commonmodel cimport CL_DEVICE_TYPE_DEFAULT, cl_get_device_id, cl_create_context
+from sunnypilot.modeld.models.commonmodel cimport mat3, ModelFrame as cppModelFrame, DrivingModelFrame as cppDrivingModelFrame, MonitoringModelFrame as cppMonitoringModelFrame
+
+
+cdef class CLContext(BaseCLContext):
+  def __cinit__(self):
+    self.device_id = cl_get_device_id(CL_DEVICE_TYPE_DEFAULT)
+    self.context = cl_create_context(self.device_id)
+
+cdef class CLMem:
+  @staticmethod
+  cdef create(void * cmem):
+    mem = CLMem()
+    mem.mem = <cl_mem*> cmem
+    return mem
+
+  @property
+  def mem_address(self):
+    return <uintptr_t>(self.mem)
+
+def cl_from_visionbuf(VisionBuf buf):
+  return CLMem.create(<void*>&buf.buf.buf_cl)
+
+
+cdef class ModelFrame:
+  cdef cppModelFrame * frame
+  cdef int buf_size
+
+  def __dealloc__(self):
+    del self.frame
+
+  def prepare(self, VisionBuf buf, float[:] projection, CLMem output):
+    cdef mat3 cprojection
+    memcpy(cprojection.v, &projection[0], 9*sizeof(float))
+    cdef unsigned char * data
+    if output is None:
+      data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, NULL)
+    else:
+      data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection, output.mem)
+    if not data:
+      return None
+
+    return np.asarray(<cnp.uint8_t[:self.buf_size]> data)
+    # return CLMem.create(data)
+
+  # def buffer_from_cl(self, CLMem in_frames):
+  #   cdef unsigned char * data2
+  #   data2 = self.frame.buffer_from_cl(in_frames.mem, self.buf_size)
+  #   return np.asarray(<cnp.uint8_t[:self.buf_size]> data2)
+
+
+cdef class DrivingModelFrame(ModelFrame):
+  cdef cppDrivingModelFrame * _frame
+
+  def __cinit__(self, CLContext context):
+    self._frame = new cppDrivingModelFrame(context.device_id, context.context)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
+
+cdef class MonitoringModelFrame(ModelFrame):
+  cdef cppMonitoringModelFrame * _frame
+
+  def __cinit__(self, CLContext context):
+    self._frame = new cppMonitoringModelFrame(context.device_id, context.context)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
--- a/sunnypilot/modeld/models/supercombo.onnx
+++ b/sunnypilot/modeld/models/supercombo.onnx
--- a/sunnypilot/modeld/models/supercombo.thneed
+++ b/sunnypilot/modeld/models/supercombo.thneed
--- a/sunnypilot/modeld/models/supercombo_metadata.pkl
+++ b/sunnypilot/modeld/models/supercombo_metadata.pkl
--- a/sunnypilot/modeld/parse_model_outputs.py
+++ b/sunnypilot/modeld/parse_model_outputs.py
@@ -0,0 +1,103 @@
+import numpy as np
+from openpilot.sunnypilot.modeld.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:
+    safe_exp(x, out=x)
+  else:
+    x = safe_exp(x)
+  x /= np.sum(x, axis=axis, keepdims=True)
+  return x
+
+class Parser:
+  def __init__(self, ignore_missing=False):
+    self.ignore_missing = ignore_missing
+
+  def check_missing(self, outs, name):
+    if name not in outs and not self.ignore_missing:
+      raise ValueError(f"Missing output {name}")
+    return name not in outs
+
+  def parse_categorical_crossentropy(self, name, outs, out_shape=None):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    if out_shape is not None:
+      raw = raw.reshape((raw.shape[0],) + out_shape)
+    outs[name] = softmax(raw, axis=-1)
+
+  def parse_binary_crossentropy(self, name, outs):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    outs[name] = sigmoid(raw)
+
+  def parse_mdn(self, name, outs, in_N=0, out_N=1, out_shape=None):
+    if self.check_missing(outs, name):
+      return
+    raw = outs[name]
+    raw = raw.reshape((raw.shape[0], max(in_N, 1), -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:
+      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)
+
+      if out_N == 1:
+        for fidx in range(weights.shape[0]):
+          idxs = np.argsort(weights[fidx][:,0])[::-1]
+          weights[fidx] = weights[fidx][idxs]
+          pred_mu[fidx] = pred_mu[fidx][idxs]
+          pred_std[fidx] = pred_std[fidx][idxs]
+      full_shape = tuple([raw.shape[0], in_N] + list(out_shape))
+      outs[name + '_weights'] = weights
+      outs[name + '_hypotheses'] = pred_mu.reshape(full_shape)
+      outs[name + '_stds_hypotheses'] = pred_std.reshape(full_shape)
+
+      pred_mu_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
+      pred_std_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
+      for fidx in range(weights.shape[0]):
+        for hidx in range(out_N):
+          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]]
+    else:
+      pred_mu_final = pred_mu
+      pred_std_final = pred_std
+
+    if out_N > 1:
+      final_shape = tuple([raw.shape[0], out_N] + list(out_shape))
+    else:
+      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,
+                   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('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('pose', outs, in_N=0, out_N=0, 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('wide_from_device_euler', outs, in_N=0, out_N=0, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
+    self.parse_mdn('lead', outs, in_N=ModelConstants.LEAD_MHP_N, out_N=ModelConstants.LEAD_MHP_SELECTION,
+                   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))
+    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))
+    return outs
--- a/sunnypilot/modeld/runners/init.py
+++ b/sunnypilot/modeld/runners/init.py
@@ -0,0 +1,27 @@
+import os
+from openpilot.system.hardware import TICI
+from openpilot.sunnypilot.modeld.runners.runmodel_pyx import RunModel, Runtime
+assert Runtime
+
+USE_THNEED = int(os.getenv('USE_THNEED', str(int(TICI))))
+USE_SNPE = int(os.getenv('USE_SNPE', str(int(TICI))))
+
+class ModelRunner(RunModel):
+  THNEED = 'THNEED'
+  SNPE = 'SNPE'
+  ONNX = 'ONNX'
+
+  def __new__(cls, paths, *args, **kwargs):
+    if ModelRunner.THNEED in paths and USE_THNEED:
+      from openpilot.sunnypilot.modeld.runners.thneedmodel_pyx import ThneedModel as Runner
+      runner_type = ModelRunner.THNEED
+    elif ModelRunner.SNPE in paths and USE_SNPE:
+      from openpilot.sunnypilot.modeld.runners.snpemodel_pyx import SNPEModel as Runner
+      runner_type = ModelRunner.SNPE
+    elif ModelRunner.ONNX in paths:
+      from openpilot.sunnypilot.modeld.runners.onnxmodel import ONNXModel as Runner
+      runner_type = ModelRunner.ONNX
+    else:
+      raise Exception("Couldn't select a model runner, make sure to pass at least one valid model path")
+
+    return Runner(str(paths[runner_type]), *args, **kwargs)
--- a/sunnypilot/modeld/runners/onnxmodel.py
+++ b/sunnypilot/modeld/runners/onnxmodel.py
@@ -0,0 +1,71 @@
+import os
+import onnx
+import sys
+import numpy as np
+from typing import Any
+
+from openpilot.sunnypilot.modeld.runners.runmodel_pyx import RunModel
+from openpilot.sunnypilot.modeld.runners.ort_helpers import convert_fp16_to_fp32, ORT_TYPES_TO_NP_TYPES
+
+
+def create_ort_session(path, fp16_to_fp32):
+  os.environ["OMP_NUM_THREADS"] = "4"
+  os.environ["OMP_WAIT_POLICY"] = "PASSIVE"
+
+  import onnxruntime as ort
+  print("Onnx available providers: ", ort.get_available_providers(), file=sys.stderr)
+  options = ort.SessionOptions()
+  options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_DISABLE_ALL
+
+  provider: str | tuple[str, dict[Any, Any]]
+  if 'OpenVINOExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
+    provider = 'OpenVINOExecutionProvider'
+  elif 'CUDAExecutionProvider' in ort.get_available_providers() and 'ONNXCPU' not in os.environ:
+    options.intra_op_num_threads = 2
+    provider = ('CUDAExecutionProvider', {'cudnn_conv_algo_search': 'EXHAUSTIVE'})
+  else:
+    options.intra_op_num_threads = 2
+    options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+    options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+    provider = 'CPUExecutionProvider'
+
+  model_data = convert_fp16_to_fp32(onnx.load(path)) if fp16_to_fp32 else path
+  print("Onnx selected provider: ", [provider], file=sys.stderr)
+  ort_session = ort.InferenceSession(model_data, options, providers=[provider])
+  print("Onnx using ", ort_session.get_providers(), file=sys.stderr)
+  return ort_session
+
+
+class ONNXModel(RunModel):
+  def __init__(self, path, output, runtime, use_tf8, cl_context):
+    self.inputs = {}
+    self.output = output
+
+    self.session = create_ort_session(path, fp16_to_fp32=True)
+    self.input_names = [x.name for x in self.session.get_inputs()]
+    self.input_shapes = {x.name: [1, *x.shape[1:]] for x in self.session.get_inputs()}
+    self.input_dtypes = {x.name: ORT_TYPES_TO_NP_TYPES[x.type] for x in self.session.get_inputs()}
+
+    # run once to initialize CUDA provider
+    if "CUDAExecutionProvider" in self.session.get_providers():
+      self.session.run(None, {k: np.zeros(self.input_shapes[k], dtype=self.input_dtypes[k]) for k in self.input_names})
+    print("ready to run onnx model", self.input_shapes, file=sys.stderr)
+
+  def addInput(self, name, buffer):
+    assert name in self.input_names
+    self.inputs[name] = buffer
+
+  def setInputBuffer(self, name, buffer):
+    assert name in self.inputs
+    self.inputs[name] = buffer
+
+  def getCLBuffer(self, name):
+    return None
+
+  def execute(self):
+    inputs = {k: v.view(self.input_dtypes[k]) for k,v in self.inputs.items()}
+    inputs = {k: v.reshape(self.input_shapes[k]).astype(self.input_dtypes[k]) for k,v in inputs.items()}
+    outputs = self.session.run(None, inputs)
+    assert len(outputs) == 1, "Only single model outputs are supported"
+    self.output[:] = outputs[0]
+    return self.output
--- a/sunnypilot/modeld/runners/ort_helpers.py
+++ b/sunnypilot/modeld/runners/ort_helpers.py
@@ -0,0 +1,36 @@
+import onnx
+import onnxruntime as ort
+import numpy as np
+import itertools
+
+ORT_TYPES_TO_NP_TYPES = {'tensor(float16)': np.float16, 'tensor(float)': np.float32, 'tensor(uint8)': np.uint8}
+
+def attributeproto_fp16_to_fp32(attr):
+  float32_list = np.frombuffer(attr.raw_data, dtype=np.float16)
+  attr.data_type = 1
+  attr.raw_data = float32_list.astype(np.float32).tobytes()
+
+def convert_fp16_to_fp32(model):
+  for i in model.graph.initializer:
+    if i.data_type == 10:
+      attributeproto_fp16_to_fp32(i)
+  for i in itertools.chain(model.graph.input, model.graph.output):
+    if i.type.tensor_type.elem_type == 10:
+      i.type.tensor_type.elem_type = 1
+  for i in model.graph.node:
+    if i.op_type == 'Cast' and i.attribute[0].i == 10:
+      i.attribute[0].i = 1
+    for a in i.attribute:
+      if hasattr(a, 't'):
+        if a.t.data_type == 10:
+          attributeproto_fp16_to_fp32(a.t)
+  return model.SerializeToString()
+
+
+def make_onnx_cpu_runner(model_path):
+  options = ort.SessionOptions()
+  options.intra_op_num_threads = 4
+  options.execution_mode = ort.ExecutionMode.ORT_SEQUENTIAL
+  options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
+  model_data = convert_fp16_to_fp32(onnx.load(model_path))
+  return ort.InferenceSession(model_data,  options, providers=['CPUExecutionProvider'])
--- a/sunnypilot/modeld/runners/run.h
+++ b/sunnypilot/modeld/runners/run.h
@@ -0,0 +1,4 @@
+#pragma once
+
+#include "sunnypilot/modeld/runners/runmodel.h"
+#include "sunnypilot/modeld/runners/snpemodel.h"
--- a/sunnypilot/modeld/runners/runmodel.h
+++ b/sunnypilot/modeld/runners/runmodel.h
@@ -0,0 +1,49 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <memory>
+#include <cassert>
+
+#include "common/clutil.h"
+#include "common/swaglog.h"
+
+#define USE_CPU_RUNTIME 0
+#define USE_GPU_RUNTIME 1
+#define USE_DSP_RUNTIME 2
+
+struct ModelInput {
+  const std::string name;
+  float *buffer;
+  int size;
+
+  ModelInput(const std::string _name, float *_buffer, int _size) : name(_name), buffer(_buffer), size(_size) {}
+  virtual void setBuffer(float *_buffer, int _size) {
+    assert(size == _size || size == 0);
+    buffer = _buffer;
+    size = _size;
+  }
+};
+
+class RunModel {
+public:
+  std::vector<std::unique_ptr<ModelInput>> inputs;
+
+  virtual ~RunModel() {}
+  virtual void execute() {}
+  virtual void* getCLBuffer(const std::string name) { return nullptr; }
+
+  virtual void addInput(const std::string name, float *buffer, int size) {
+    inputs.push_back(std::unique_ptr<ModelInput>(new ModelInput(name, buffer, size)));
+  }
+  virtual void setInputBuffer(const std::string name, float *buffer, int size) {
+    for (auto &input : inputs) {
+      if (name == input->name) {
+        input->setBuffer(buffer, size);
+        return;
+      }
+    }
+    LOGE("Tried to update input `%s` but no input with this name exists", name.c_str());
+    assert(false);
+  }
+};
--- a/sunnypilot/modeld/runners/runmodel.pxd
+++ b/sunnypilot/modeld/runners/runmodel.pxd
@@ -0,0 +1,14 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+cdef extern from "sunnypilot/modeld/runners/runmodel.h":
+  cdef int USE_CPU_RUNTIME
+  cdef int USE_GPU_RUNTIME
+  cdef int USE_DSP_RUNTIME
+
+  cdef cppclass RunModel:
+    void addInput(string, float*, int)
+    void setInputBuffer(string, float*, int)
+    void * getCLBuffer(string)
+    void execute()
--- a/sunnypilot/modeld/runners/runmodel_pyx.pxd
+++ b/sunnypilot/modeld/runners/runmodel_pyx.pxd
@@ -0,0 +1,6 @@
+# distutils: language = c++
+
+from .runmodel cimport RunModel as cppRunModel
+
+cdef class RunModel:
+  cdef cppRunModel * model
--- a/sunnypilot/modeld/runners/runmodel_pyx.pyx
+++ b/sunnypilot/modeld/runners/runmodel_pyx.pyx
@@ -0,0 +1,37 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+from libcpp.string cimport string
+
+from .runmodel cimport USE_CPU_RUNTIME, USE_GPU_RUNTIME, USE_DSP_RUNTIME
+from sunnypilot.modeld.models.commonmodel_pyx cimport CLMem
+
+class Runtime:
+  CPU = USE_CPU_RUNTIME
+  GPU = USE_GPU_RUNTIME
+  DSP = USE_DSP_RUNTIME
+
+cdef class RunModel:
+  def __dealloc__(self):
+    del self.model
+
+  def addInput(self, string name, float[:] buffer):
+    if buffer is not None:
+      self.model.addInput(name, &buffer[0], len(buffer))
+    else:
+      self.model.addInput(name, NULL, 0)
+
+  def setInputBuffer(self, string name, float[:] buffer):
+    if buffer is not None:
+      self.model.setInputBuffer(name, &buffer[0], len(buffer))
+    else:
+      self.model.setInputBuffer(name, NULL, 0)
+
+  def getCLBuffer(self, string name):
+    cdef void * cl_buf = self.model.getCLBuffer(name)
+    if not cl_buf:
+      return None
+    return CLMem.create(cl_buf)
+
+  def execute(self):
+    self.model.execute()
--- a/sunnypilot/modeld/runners/snpemodel.cc
+++ b/sunnypilot/modeld/runners/snpemodel.cc
@@ -0,0 +1,116 @@
+#pragma clang diagnostic ignored "-Wexceptions"
+
+#include "sunnypilot/modeld/runners/snpemodel.h"
+
+#include <cstring>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common/util.h"
+#include "common/timing.h"
+
+void PrintErrorStringAndExit() {
+  std::cerr << zdl::DlSystem::getLastErrorString() << std::endl;
+  std::exit(EXIT_FAILURE);
+}
+
+SNPEModel::SNPEModel(const std::string path, float *_output, size_t _output_size, int runtime, bool _use_tf8, cl_context context) {
+  output = _output;
+  output_size = _output_size;
+  use_tf8 = _use_tf8;
+
+#ifdef QCOM2
+  if (runtime == USE_GPU_RUNTIME) {
+    snpe_runtime = zdl::DlSystem::Runtime_t::GPU;
+  } else if (runtime == USE_DSP_RUNTIME) {
+    snpe_runtime = zdl::DlSystem::Runtime_t::DSP;
+  } else {
+    snpe_runtime = zdl::DlSystem::Runtime_t::CPU;
+  }
+  assert(zdl::SNPE::SNPEFactory::isRuntimeAvailable(snpe_runtime));
+#endif
+  model_data = util::read_file(path);
+  assert(model_data.size() > 0);
+
+  // load model
+  std::unique_ptr<zdl::DlContainer::IDlContainer> container = zdl::DlContainer::IDlContainer::open((uint8_t*)model_data.data(), model_data.size());
+  if (!container) { PrintErrorStringAndExit(); }
+  LOGW("loaded model with size: %lu", model_data.size());
+
+  // create model runner
+  zdl::SNPE::SNPEBuilder snpe_builder(container.get());
+  while (!snpe) {
+#ifdef QCOM2
+    snpe = snpe_builder.setOutputLayers({})
+                       .setRuntimeProcessor(snpe_runtime)
+                       .setUseUserSuppliedBuffers(true)
+                       .setPerformanceProfile(zdl::DlSystem::PerformanceProfile_t::HIGH_PERFORMANCE)
+                       .build();
+#else
+    snpe = snpe_builder.setOutputLayers({})
+                       .setUseUserSuppliedBuffers(true)
+                       .setPerformanceProfile(zdl::DlSystem::PerformanceProfile_t::HIGH_PERFORMANCE)
+                       .build();
+#endif
+    if (!snpe) std::cerr << zdl::DlSystem::getLastErrorString() << std::endl;
+  }
+
+  // create output buffer
+  zdl::DlSystem::UserBufferEncodingFloat ub_encoding_float;
+  zdl::DlSystem::IUserBufferFactory &ub_factory = zdl::SNPE::SNPEFactory::getUserBufferFactory();
+
+  const auto &output_tensor_names_opt = snpe->getOutputTensorNames();
+  if (!output_tensor_names_opt) throw std::runtime_error("Error obtaining output tensor names");
+  const auto &output_tensor_names = *output_tensor_names_opt;
+  assert(output_tensor_names.size() == 1);
+  const char *output_tensor_name = output_tensor_names.at(0);
+  const zdl::DlSystem::TensorShape &buffer_shape = snpe->getInputOutputBufferAttributes(output_tensor_name)->getDims();
+  if (output_size != 0) {
+    assert(output_size == buffer_shape[1]);
+  } else {
+    output_size = buffer_shape[1];
+  }
+  std::vector<size_t> output_strides = {output_size * sizeof(float), sizeof(float)};
+  output_buffer = ub_factory.createUserBuffer(output, output_size * sizeof(float), output_strides, &ub_encoding_float);
+  output_map.add(output_tensor_name, output_buffer.get());
+}
+
+void SNPEModel::addInput(const std::string name, float *buffer, int size) {
+  const int idx = inputs.size();
+  const auto &input_tensor_names_opt = snpe->getInputTensorNames();
+  if (!input_tensor_names_opt) throw std::runtime_error("Error obtaining input tensor names");
+  const auto &input_tensor_names = *input_tensor_names_opt;
+  const char *input_tensor_name = input_tensor_names.at(idx);
+  const bool input_tf8 = use_tf8 && strcmp(input_tensor_name, "input_img") == 0;  // TODO: This is a terrible hack, get rid of this name check both here and in onnx_runner.py
+  LOGW("adding index %d: %s", idx, input_tensor_name);
+
+  zdl::DlSystem::UserBufferEncodingFloat ub_encoding_float;
+  zdl::DlSystem::UserBufferEncodingTf8 ub_encoding_tf8(0, 1./255); // network takes 0-1
+  zdl::DlSystem::IUserBufferFactory &ub_factory = zdl::SNPE::SNPEFactory::getUserBufferFactory();
+  zdl::DlSystem::UserBufferEncoding *input_encoding = input_tf8 ? (zdl::DlSystem::UserBufferEncoding*)&ub_encoding_tf8 : (zdl::DlSystem::UserBufferEncoding*)&ub_encoding_float;
+
+  const auto &buffer_shape_opt = snpe->getInputDimensions(input_tensor_name);
+  const zdl::DlSystem::TensorShape &buffer_shape = *buffer_shape_opt;
+  size_t size_of_input = input_tf8 ? sizeof(uint8_t) : sizeof(float);
+  std::vector<size_t> strides(buffer_shape.rank());
+  strides[strides.size() - 1] = size_of_input;
+  size_t product = 1;
+  for (size_t i = 0; i < buffer_shape.rank(); i++) product *= buffer_shape[i];
+  size_t stride = strides[strides.size() - 1];
+  for (size_t i = buffer_shape.rank() - 1; i > 0; i--) {
+    stride *= buffer_shape[i];
+    strides[i-1] = stride;
+  }
+
+  auto input_buffer = ub_factory.createUserBuffer(buffer, product*size_of_input, strides, input_encoding);
+  input_map.add(input_tensor_name, input_buffer.get());
+  inputs.push_back(std::unique_ptr<SNPEModelInput>(new SNPEModelInput(name, buffer, size, std::move(input_buffer))));
+}
+
+void SNPEModel::execute() {
+  if (!snpe->execute(input_map, output_map)) {
+    PrintErrorStringAndExit();
+  }
+}
--- a/sunnypilot/modeld/runners/snpemodel.h
+++ b/sunnypilot/modeld/runners/snpemodel.h
@@ -0,0 +1,52 @@
+#pragma once
+#pragma clang diagnostic ignored "-Wdeprecated-declarations"
+
+#include <memory>
+#include <string>
+#include <utility>
+
+#include <DlContainer/IDlContainer.hpp>
+#include <DlSystem/DlError.hpp>
+#include <DlSystem/ITensor.hpp>
+#include <DlSystem/ITensorFactory.hpp>
+#include <DlSystem/IUserBuffer.hpp>
+#include <DlSystem/IUserBufferFactory.hpp>
+#include <SNPE/SNPE.hpp>
+#include <SNPE/SNPEBuilder.hpp>
+#include <SNPE/SNPEFactory.hpp>
+
+#include "sunnypilot/modeld/runners/runmodel.h"
+
+struct SNPEModelInput : public ModelInput {
+  std::unique_ptr<zdl::DlSystem::IUserBuffer> snpe_buffer;
+
+  SNPEModelInput(const std::string _name, float *_buffer, int _size, std::unique_ptr<zdl::DlSystem::IUserBuffer> _snpe_buffer) : ModelInput(_name, _buffer, _size), snpe_buffer(std::move(_snpe_buffer)) {}
+  void setBuffer(float *_buffer, int _size) {
+    ModelInput::setBuffer(_buffer, _size);
+    assert(snpe_buffer->setBufferAddress(_buffer) == true);
+  }
+};
+
+class SNPEModel : public RunModel {
+public:
+  SNPEModel(const std::string path, float *_output, size_t _output_size, int runtime, bool use_tf8 = false, cl_context context = NULL);
+  void addInput(const std::string name, float *buffer, int size);
+  void execute();
+
+private:
+  std::string model_data;
+
+#ifdef QCOM2
+  zdl::DlSystem::Runtime_t snpe_runtime;
+#endif
+
+  // snpe model stuff
+  std::unique_ptr<zdl::SNPE::SNPE> snpe;
+  zdl::DlSystem::UserBufferMap input_map;
+  zdl::DlSystem::UserBufferMap output_map;
+  std::unique_ptr<zdl::DlSystem::IUserBuffer> output_buffer;
+
+  bool use_tf8;
+  float *output;
+  size_t output_size;
+};
--- a/sunnypilot/modeld/runners/snpemodel.pxd
+++ b/sunnypilot/modeld/runners/snpemodel.pxd
@@ -0,0 +1,9 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+from msgq.visionipc.visionipc cimport cl_context
+
+cdef extern from "sunnypilot/modeld/runners/snpemodel.h":
+  cdef cppclass SNPEModel:
+    SNPEModel(string, float*, size_t, int, bool, cl_context)
--- a/sunnypilot/modeld/runners/snpemodel_pyx.pyx
+++ b/sunnypilot/modeld/runners/snpemodel_pyx.pyx
@@ -0,0 +1,17 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+import os
+from libcpp cimport bool
+from libcpp.string cimport string
+
+from .snpemodel cimport SNPEModel as cppSNPEModel
+from sunnypilot.modeld.models.commonmodel_pyx cimport CLContext
+from sunnypilot.modeld.runners.runmodel_pyx cimport RunModel
+from sunnypilot.modeld.runners.runmodel cimport RunModel as cppRunModel
+
+os.environ['ADSP_LIBRARY_PATH'] = "/data/pythonpath/third_party/snpe/dsp/"
+
+cdef class SNPEModel(RunModel):
+  def __cinit__(self, string path, float[:] output, int runtime, bool use_tf8, CLContext context):
+    self.model = <cppRunModel *> new cppSNPEModel(path, &output[0], len(output), runtime, use_tf8, context.context)
--- a/sunnypilot/modeld/runners/thneedmodel.cc
+++ b/sunnypilot/modeld/runners/thneedmodel.cc
@@ -0,0 +1,58 @@
+#include "sunnypilot/modeld/runners/thneedmodel.h"
+
+#include <string>
+
+#include "common/swaglog.h"
+
+ThneedModel::ThneedModel(const std::string path, float *_output, size_t _output_size, int runtime, bool luse_tf8, cl_context context) {
+  thneed = new Thneed(true, context);
+  thneed->load(path.c_str());
+  thneed->clexec();
+
+  recorded = false;
+  output = _output;
+}
+
+void* ThneedModel::getCLBuffer(const std::string name) {
+  int index = -1;
+  for (int i = 0; i < inputs.size(); i++) {
+    if (name == inputs[i]->name) {
+      index = i;
+      break;
+    }
+  }
+
+  if (index == -1) {
+    LOGE("Tried to get CL buffer for input `%s` but no input with this name exists", name.c_str());
+    assert(false);
+  }
+
+  if (thneed->input_clmem.size() >= inputs.size()) {
+    return &thneed->input_clmem[inputs.size() - index - 1];
+  } else {
+    return nullptr;
+  }
+}
+
+void ThneedModel::execute() {
+  if (!recorded) {
+    thneed->record = true;
+    float *input_buffers[inputs.size()];
+    for (int i = 0; i < inputs.size(); i++) {
+      input_buffers[inputs.size() - i - 1] = inputs[i]->buffer;
+    }
+
+    thneed->copy_inputs(input_buffers);
+    thneed->clexec();
+    thneed->copy_output(output);
+    thneed->stop();
+
+    recorded = true;
+  } else {
+    float *input_buffers[inputs.size()];
+    for (int i = 0; i < inputs.size(); i++) {
+      input_buffers[inputs.size() - i - 1] = inputs[i]->buffer;
+    }
+    thneed->execute(input_buffers, output);
+  }
+}
--- a/sunnypilot/modeld/runners/thneedmodel.h
+++ b/sunnypilot/modeld/runners/thneedmodel.h
@@ -0,0 +1,17 @@
+#pragma once
+
+#include <string>
+
+#include "sunnypilot/modeld/runners/runmodel.h"
+#include "sunnypilot/modeld/thneed/thneed.h"
+
+class ThneedModel : public RunModel {
+public:
+  ThneedModel(const std::string path, float *_output, size_t _output_size, int runtime, bool use_tf8 = false, cl_context context = NULL);
+  void *getCLBuffer(const std::string name);
+  void execute();
+private:
+  Thneed *thneed = NULL;
+  bool recorded;
+  float *output;
+};
--- a/sunnypilot/modeld/runners/thneedmodel.pxd
+++ b/sunnypilot/modeld/runners/thneedmodel.pxd
@@ -0,0 +1,9 @@
+# distutils: language = c++
+
+from libcpp.string cimport string
+
+from msgq.visionipc.visionipc cimport cl_context
+
+cdef extern from "sunnypilot/modeld/runners/thneedmodel.h":
+  cdef cppclass ThneedModel:
+    ThneedModel(string, float*, size_t, int, bool, cl_context)
--- a/sunnypilot/modeld/runners/thneedmodel_pyx.pyx
+++ b/sunnypilot/modeld/runners/thneedmodel_pyx.pyx
@@ -0,0 +1,14 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+from libcpp cimport bool
+from libcpp.string cimport string
+
+from .thneedmodel cimport ThneedModel as cppThneedModel
+from sunnypilot.modeld.models.commonmodel_pyx cimport CLContext
+from sunnypilot.modeld.runners.runmodel_pyx cimport RunModel
+from sunnypilot.modeld.runners.runmodel cimport RunModel as cppRunModel
+
+cdef class ThneedModel(RunModel):
+  def __cinit__(self, string path, float[:] output, int runtime, bool use_tf8, CLContext context):
+    self.model = <cppRunModel *> new cppThneedModel(path, &output[0], len(output), runtime, use_tf8, context.context)
--- a/sunnypilot/modeld/thneed/README
+++ b/sunnypilot/modeld/thneed/README
@@ -0,0 +1,8 @@
+thneed is an SNPE accelerator. I know SNPE is already an accelerator, but sometimes things need to go even faster..
+
+It runs on the local device, and caches a single model run. Then it replays it, but fast.
+
+thneed slices through abstraction layers like a fish.
+
+You need a thneed.
+
--- a/sunnypilot/modeld/thneed/init.py
+++ b/sunnypilot/modeld/thneed/init.py
--- a/sunnypilot/modeld/thneed/serialize.cc
+++ b/sunnypilot/modeld/thneed/serialize.cc
@@ -0,0 +1,154 @@
+#include <cassert>
+#include <set>
+
+#include "third_party/json11/json11.hpp"
+#include "common/util.h"
+#include "common/clutil.h"
+#include "common/swaglog.h"
+#include "sunnypilot/modeld/thneed/thneed.h"
+using namespace json11;
+
+extern map<cl_program, string> g_program_source;
+
+void Thneed::load(const char *filename) {
+  LOGD("Thneed::load: loading from %s\n", filename);
+
+  string buf = util::read_file(filename);
+  int jsz = *(int *)buf.data();
+  string jsonerr;
+  string jj(buf.data() + sizeof(int), jsz);
+  Json jdat = Json::parse(jj, jsonerr);
+
+  map<cl_mem, cl_mem> real_mem;
+  real_mem[NULL] = NULL;
+
+  int ptr = sizeof(int)+jsz;
+  for (auto &obj : jdat["objects"].array_items()) {
+    auto mobj = obj.object_items();
+    int sz = mobj["size"].int_value();
+    cl_mem clbuf = NULL;
+
+    if (mobj["buffer_id"].string_value().size() > 0) {
+      // image buffer must already be allocated
+      clbuf = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())];
+      assert(mobj["needs_load"].bool_value() == false);
+    } else {
+      if (mobj["needs_load"].bool_value()) {
+        clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, &buf[ptr], NULL);
+        if (debug >= 1) printf("loading %p %d @ 0x%X\n", clbuf, sz, ptr);
+        ptr += sz;
+      } else {
+        // TODO: is there a faster way to init zeroed out buffers?
+        void *host_zeros = calloc(sz, 1);
+        clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, host_zeros, NULL);
+        free(host_zeros);
+      }
+    }
+    assert(clbuf != NULL);
+
+    if (mobj["arg_type"] == "image2d_t" || mobj["arg_type"] == "image1d_t") {
+      cl_image_desc desc = {0};
+      desc.image_type = (mobj["arg_type"] == "image2d_t") ? CL_MEM_OBJECT_IMAGE2D : CL_MEM_OBJECT_IMAGE1D_BUFFER;
+      desc.image_width = mobj["width"].int_value();
+      desc.image_height = mobj["height"].int_value();
+      desc.image_row_pitch = mobj["row_pitch"].int_value();
+      assert(sz == desc.image_height*desc.image_row_pitch);
+#ifdef QCOM2
+      desc.buffer = clbuf;
+#else
+      // TODO: we are creating unused buffers on PC
+      clReleaseMemObject(clbuf);
+#endif
+      cl_image_format format = {0};
+      format.image_channel_order = CL_RGBA;
+      format.image_channel_data_type = mobj["float32"].bool_value() ? CL_FLOAT : CL_HALF_FLOAT;
+
+      cl_int errcode;
+
+#ifndef QCOM2
+      if (mobj["needs_load"].bool_value()) {
+        clbuf = clCreateImage(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, &format, &desc, &buf[ptr-sz], &errcode);
+      } else {
+        clbuf = clCreateImage(context, CL_MEM_READ_WRITE, &format, &desc, NULL, &errcode);
+      }
+#else
+      clbuf = clCreateImage(context, CL_MEM_READ_WRITE, &format, &desc, NULL, &errcode);
+#endif
+      if (clbuf == NULL) {
+        LOGE("clError: %s create image %zux%zu rp %zu with buffer %p\n", cl_get_error_string(errcode),
+             desc.image_width, desc.image_height, desc.image_row_pitch, desc.buffer);
+      }
+      assert(clbuf != NULL);
+    }
+
+    real_mem[*(cl_mem*)(mobj["id"].string_value().data())] = clbuf;
+  }
+
+  map<string, cl_program> g_programs;
+  for (const auto &[name, source] : jdat["programs"].object_items()) {
+    if (debug >= 1) printf("building %s with size %zu\n", name.c_str(), source.string_value().size());
+    g_programs[name] = cl_program_from_source(context, device_id, source.string_value());
+  }
+
+  for (auto &obj : jdat["inputs"].array_items()) {
+    auto mobj = obj.object_items();
+    int sz = mobj["size"].int_value();
+    cl_mem aa = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())];
+    input_clmem.push_back(aa);
+    input_sizes.push_back(sz);
+    LOGD("Thneed::load: adding input %s with size %d\n", mobj["name"].string_value().data(), sz);
+
+    cl_int cl_err;
+    void *ret = clEnqueueMapBuffer(command_queue, aa, CL_TRUE, CL_MAP_WRITE, 0, sz, 0, NULL, NULL, &cl_err);
+    if (cl_err != CL_SUCCESS) LOGE("clError: %s map %p %d\n", cl_get_error_string(cl_err), aa, sz);
+    assert(cl_err == CL_SUCCESS);
+    inputs.push_back(ret);
+  }
+
+  for (auto &obj : jdat["outputs"].array_items()) {
+    auto mobj = obj.object_items();
+    int sz = mobj["size"].int_value();
+    LOGD("Thneed::save: adding output with size %d\n", sz);
+    // TODO: support multiple outputs
+    output = real_mem[*(cl_mem*)(mobj["buffer_id"].string_value().data())];
+    assert(output != NULL);
+  }
+
+  for (auto &obj : jdat["binaries"].array_items()) {
+    string name = obj["name"].string_value();
+    size_t length = obj["length"].int_value();
+    if (debug >= 1) printf("binary %s with size %zu\n", name.c_str(), length);
+    g_programs[name] = cl_program_from_binary(context, device_id, (const uint8_t*)&buf[ptr], length);
+    ptr += length;
+  }
+
+  for (auto &obj : jdat["kernels"].array_items()) {
+    auto gws = obj["global_work_size"];
+    auto lws = obj["local_work_size"];
+    auto kk = shared_ptr<CLQueuedKernel>(new CLQueuedKernel(this));
+
+    kk->name = obj["name"].string_value();
+    kk->program = g_programs[kk->name];
+    kk->work_dim = obj["work_dim"].int_value();
+    for (int i = 0; i < kk->work_dim; i++) {
+      kk->global_work_size[i] = gws[i].int_value();
+      kk->local_work_size[i] = lws[i].int_value();
+    }
+    kk->num_args = obj["num_args"].int_value();
+    for (int i = 0; i < kk->num_args; i++) {
+      string arg = obj["args"].array_items()[i].string_value();
+      int arg_size = obj["args_size"].array_items()[i].int_value();
+      kk->args_size.push_back(arg_size);
+      if (arg_size == 8) {
+        cl_mem val = *(cl_mem*)(arg.data());
+        val = real_mem[val];
+        kk->args.push_back(string((char*)&val, sizeof(val)));
+      } else {
+        kk->args.push_back(arg);
+      }
+    }
+    kq.push_back(kk);
+  }
+
+  clFinish(command_queue);
+}
--- a/sunnypilot/modeld/thneed/thneed.h
+++ b/sunnypilot/modeld/thneed/thneed.h
@@ -0,0 +1,133 @@
+#pragma once
+
+#ifndef __user
+#define __user __attribute__(())
+#endif
+
+#include <cstdint>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include <CL/cl.h>
+
+#include "third_party/linux/include/msm_kgsl.h"
+
+using namespace std;
+
+cl_int thneed_clSetKernelArg(cl_kernel kernel, cl_uint arg_index, size_t arg_size, const void *arg_value);
+
+namespace json11 {
+  class Json;
+}
+class Thneed;
+
+class GPUMalloc {
+  public:
+    GPUMalloc(int size, int fd);
+    ~GPUMalloc();
+    void *alloc(int size);
+  private:
+    uint64_t base;
+    int remaining;
+};
+
+class CLQueuedKernel {
+  public:
+    CLQueuedKernel(Thneed *lthneed) { thneed = lthneed; }
+    CLQueuedKernel(Thneed *lthneed,
+                   cl_kernel _kernel,
+                   cl_uint _work_dim,
+                   const size_t *_global_work_size,
+                   const size_t *_local_work_size);
+    cl_int exec();
+    void debug_print(bool verbose);
+    int get_arg_num(const char *search_arg_name);
+    cl_program program;
+    string name;
+    cl_uint num_args;
+    vector<string> arg_names;
+    vector<string> arg_types;
+    vector<string> args;
+    vector<int> args_size;
+    cl_kernel kernel = NULL;
+    json11::Json to_json() const;
+
+    cl_uint work_dim;
+    size_t global_work_size[3] = {0};
+    size_t local_work_size[3] = {0};
+  private:
+    Thneed *thneed;
+};
+
+class CachedIoctl {
+  public:
+    virtual void exec() {}
+};
+
+class CachedSync: public CachedIoctl {
+  public:
+    CachedSync(Thneed *lthneed, string ldata) { thneed = lthneed; data = ldata; }
+    void exec();
+  private:
+    Thneed *thneed;
+    string data;
+};
+
+class CachedCommand: public CachedIoctl {
+  public:
+    CachedCommand(Thneed *lthneed, struct kgsl_gpu_command *cmd);
+    void exec();
+  private:
+    void disassemble(int cmd_index);
+    struct kgsl_gpu_command cache;
+    unique_ptr<kgsl_command_object[]> cmds;
+    unique_ptr<kgsl_command_object[]> objs;
+    Thneed *thneed;
+    vector<shared_ptr<CLQueuedKernel> > kq;
+};
+
+class Thneed {
+  public:
+    Thneed(bool do_clinit=false, cl_context _context = NULL);
+    void stop();
+    void execute(float **finputs, float *foutput, bool slow=false);
+    void wait();
+
+    vector<cl_mem> input_clmem;
+    vector<void *> inputs;
+    vector<size_t> input_sizes;
+    cl_mem output = NULL;
+
+    cl_context context = NULL;
+    cl_command_queue command_queue;
+    cl_device_id device_id;
+    int context_id;
+
+    // protected?
+    bool record = false;
+    int debug;
+    int timestamp;
+
+#ifdef QCOM2
+    unique_ptr<GPUMalloc> ram;
+    vector<unique_ptr<CachedIoctl> > cmds;
+    int fd;
+#endif
+
+    // all CL kernels
+    void copy_inputs(float **finputs, bool internal=false);
+    void copy_output(float *foutput);
+    cl_int clexec();
+    vector<shared_ptr<CLQueuedKernel> > kq;
+
+    // pending CL kernels
+    vector<shared_ptr<CLQueuedKernel> > ckq;
+
+    // loading
+    void load(const char *filename);
+  private:
+    void clinit();
+};
+
--- a/sunnypilot/modeld/thneed/thneed_common.cc
+++ b/sunnypilot/modeld/thneed/thneed_common.cc
@@ -0,0 +1,216 @@
+#include "sunnypilot/modeld/thneed/thneed.h"
+
+#include <cassert>
+#include <cstring>
+#include <map>
+
+#include "common/clutil.h"
+#include "common/timing.h"
+
+map<pair<cl_kernel, int>, string> g_args;
+map<pair<cl_kernel, int>, int> g_args_size;
+map<cl_program, string> g_program_source;
+
+void Thneed::stop() {
+  //printf("Thneed::stop: recorded %lu commands\n", cmds.size());
+  record = false;
+}
+
+void Thneed::clinit() {
+  device_id = cl_get_device_id(CL_DEVICE_TYPE_DEFAULT);
+  if (context == NULL) context = CL_CHECK_ERR(clCreateContext(NULL, 1, &device_id, NULL, NULL, &err));
+  //cl_command_queue_properties props[3] = {CL_QUEUE_PROPERTIES, CL_QUEUE_PROFILING_ENABLE, 0};
+  cl_command_queue_properties props[3] = {CL_QUEUE_PROPERTIES, 0, 0};
+  command_queue = CL_CHECK_ERR(clCreateCommandQueueWithProperties(context, device_id, props, &err));
+  printf("Thneed::clinit done\n");
+}
+
+cl_int Thneed::clexec() {
+  if (debug >= 1) printf("Thneed::clexec: running %lu queued kernels\n", kq.size());
+  for (auto &k : kq) {
+    if (record) ckq.push_back(k);
+    cl_int ret = k->exec();
+    assert(ret == CL_SUCCESS);
+  }
+  return clFinish(command_queue);
+}
+
+void Thneed::copy_inputs(float **finputs, bool internal) {
+  for (int idx = 0; idx < inputs.size(); ++idx) {
+    if (debug >= 1) printf("copying %lu -- %p -> %p (cl %p)\n", input_sizes[idx], finputs[idx], inputs[idx], input_clmem[idx]);
+
+    if (internal) {
+      // if it's internal, using memcpy is fine since the buffer sync is cached in the ioctl layer
+      if (finputs[idx] != NULL) memcpy(inputs[idx], finputs[idx], input_sizes[idx]);
+    } else {
+      if (finputs[idx] != NULL) CL_CHECK(clEnqueueWriteBuffer(command_queue, input_clmem[idx], CL_TRUE, 0, input_sizes[idx], finputs[idx], 0, NULL, NULL));
+    }
+  }
+}
+
+void Thneed::copy_output(float *foutput) {
+  if (output != NULL) {
+    size_t sz;
+    clGetMemObjectInfo(output, CL_MEM_SIZE, sizeof(sz), &sz, NULL);
+    if (debug >= 1) printf("copying %lu for output %p -> %p\n", sz, output, foutput);
+    CL_CHECK(clEnqueueReadBuffer(command_queue, output, CL_TRUE, 0, sz, foutput, 0, NULL, NULL));
+  } else {
+    printf("CAUTION: model output is NULL, does it have no outputs?\n");
+  }
+}
+
+// *********** CLQueuedKernel ***********
+
+CLQueuedKernel::CLQueuedKernel(Thneed *lthneed,
+                               cl_kernel _kernel,
+                               cl_uint _work_dim,
+                               const size_t *_global_work_size,
+                               const size_t *_local_work_size) {
+  thneed = lthneed;
+  kernel = _kernel;
+  work_dim = _work_dim;
+  assert(work_dim <= 3);
+  for (int i = 0; i < work_dim; i++) {
+    global_work_size[i] = _global_work_size[i];
+    local_work_size[i] = _local_work_size[i];
+  }
+
+  char _name[0x100];
+  clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME, sizeof(_name), _name, NULL);
+  name = string(_name);
+  clGetKernelInfo(kernel, CL_KERNEL_NUM_ARGS, sizeof(num_args), &num_args, NULL);
+
+  // get args
+  for (int i = 0; i < num_args; i++) {
+    char arg_name[0x100] = {0};
+    clGetKernelArgInfo(kernel, i, CL_KERNEL_ARG_NAME, sizeof(arg_name), arg_name, NULL);
+    arg_names.push_back(string(arg_name));
+    clGetKernelArgInfo(kernel, i, CL_KERNEL_ARG_TYPE_NAME, sizeof(arg_name), arg_name, NULL);
+    arg_types.push_back(string(arg_name));
+
+    args.push_back(g_args[make_pair(kernel, i)]);
+    args_size.push_back(g_args_size[make_pair(kernel, i)]);
+  }
+
+  // get program
+  clGetKernelInfo(kernel, CL_KERNEL_PROGRAM, sizeof(program), &program, NULL);
+}
+
+int CLQueuedKernel::get_arg_num(const char *search_arg_name) {
+  for (int i = 0; i < num_args; i++) {
+    if (arg_names[i] == search_arg_name) return i;
+  }
+  printf("failed to find %s in %s\n", search_arg_name, name.c_str());
+  assert(false);
+}
+
+cl_int CLQueuedKernel::exec() {
+  if (kernel == NULL) {
+    kernel = clCreateKernel(program, name.c_str(), NULL);
+    arg_names.clear();
+    arg_types.clear();
+
+    for (int j = 0; j < num_args; j++) {
+      char arg_name[0x100] = {0};
+      clGetKernelArgInfo(kernel, j, CL_KERNEL_ARG_NAME, sizeof(arg_name), arg_name, NULL);
+      arg_names.push_back(string(arg_name));
+      clGetKernelArgInfo(kernel, j, CL_KERNEL_ARG_TYPE_NAME, sizeof(arg_name), arg_name, NULL);
+      arg_types.push_back(string(arg_name));
+
+      cl_int ret;
+      if (args[j].size() != 0) {
+        assert(args[j].size() == args_size[j]);
+        ret = thneed_clSetKernelArg(kernel, j, args[j].size(), args[j].data());
+      } else {
+        ret = thneed_clSetKernelArg(kernel, j, args_size[j], NULL);
+      }
+      assert(ret == CL_SUCCESS);
+    }
+  }
+
+  if (thneed->debug >= 1) {
+    debug_print(thneed->debug >= 2);
+  }
+
+  return clEnqueueNDRangeKernel(thneed->command_queue,
+    kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, NULL);
+}
+
+void CLQueuedKernel::debug_print(bool verbose) {
+  printf("%p %56s -- ", kernel, name.c_str());
+  for (int i = 0; i < work_dim; i++) {
+    printf("%4zu ", global_work_size[i]);
+  }
+  printf(" -- ");
+  for (int i = 0; i < work_dim; i++) {
+    printf("%4zu ", local_work_size[i]);
+  }
+  printf("\n");
+
+  if (verbose) {
+    for (int i = 0; i < num_args; i++) {
+      string arg = args[i];
+      printf("  %s %s", arg_types[i].c_str(), arg_names[i].c_str());
+      void *arg_value = (void*)arg.data();
+      int arg_size = arg.size();
+      if (arg_size == 0) {
+        printf(" (size) %d", args_size[i]);
+      } else if (arg_size == 1) {
+        printf(" = %d", *((char*)arg_value));
+      } else if (arg_size == 2) {
+        printf(" = %d", *((short*)arg_value));
+      } else if (arg_size == 4) {
+        if (arg_types[i] == "float") {
+          printf(" = %f", *((float*)arg_value));
+        } else {
+          printf(" = %d", *((int*)arg_value));
+        }
+      } else if (arg_size == 8) {
+        cl_mem val = (cl_mem)(*((uintptr_t*)arg_value));
+        printf(" = %p", val);
+        if (val != NULL) {
+          cl_mem_object_type obj_type;
+          clGetMemObjectInfo(val, CL_MEM_TYPE, sizeof(obj_type), &obj_type, NULL);
+          if (arg_types[i] == "image2d_t" || arg_types[i] == "image1d_t" || obj_type == CL_MEM_OBJECT_IMAGE2D) {
+            cl_image_format format;
+            size_t width, height, depth, array_size, row_pitch, slice_pitch;
+            cl_mem buf;
+            clGetImageInfo(val, CL_IMAGE_FORMAT, sizeof(format), &format, NULL);
+            assert(format.image_channel_order == CL_RGBA);
+            assert(format.image_channel_data_type == CL_HALF_FLOAT || format.image_channel_data_type == CL_FLOAT);
+            clGetImageInfo(val, CL_IMAGE_WIDTH, sizeof(width), &width, NULL);
+            clGetImageInfo(val, CL_IMAGE_HEIGHT, sizeof(height), &height, NULL);
+            clGetImageInfo(val, CL_IMAGE_ROW_PITCH, sizeof(row_pitch), &row_pitch, NULL);
+            clGetImageInfo(val, CL_IMAGE_DEPTH, sizeof(depth), &depth, NULL);
+            clGetImageInfo(val, CL_IMAGE_ARRAY_SIZE, sizeof(array_size), &array_size, NULL);
+            clGetImageInfo(val, CL_IMAGE_SLICE_PITCH, sizeof(slice_pitch), &slice_pitch, NULL);
+            assert(depth == 0);
+            assert(array_size == 0);
+            assert(slice_pitch == 0);
+
+            clGetImageInfo(val, CL_IMAGE_BUFFER, sizeof(buf), &buf, NULL);
+            size_t sz = 0;
+            if (buf != NULL) clGetMemObjectInfo(buf, CL_MEM_SIZE, sizeof(sz), &sz, NULL);
+            printf(" image %zu x %zu rp %zu @ %p buffer %zu", width, height, row_pitch, buf, sz);
+          } else {
+            size_t sz;
+            clGetMemObjectInfo(val, CL_MEM_SIZE, sizeof(sz), &sz, NULL);
+            printf(" buffer %zu", sz);
+          }
+        }
+      }
+      printf("\n");
+    }
+  }
+}
+
+cl_int thneed_clSetKernelArg(cl_kernel kernel, cl_uint arg_index, size_t arg_size, const void *arg_value) {
+  g_args_size[make_pair(kernel, arg_index)] = arg_size;
+  if (arg_value != NULL) {
+    g_args[make_pair(kernel, arg_index)] = string((char*)arg_value, arg_size);
+  } else {
+    g_args[make_pair(kernel, arg_index)] = string("");
+  }
+  cl_int ret = clSetKernelArg(kernel, arg_index, arg_size, arg_value);
+  return ret;
+}
--- a/sunnypilot/modeld/thneed/thneed_pc.cc
+++ b/sunnypilot/modeld/thneed/thneed_pc.cc
@@ -0,0 +1,32 @@
+#include "sunnypilot/modeld/thneed/thneed.h"
+
+#include <cassert>
+
+#include "common/clutil.h"
+#include "common/timing.h"
+
+Thneed::Thneed(bool do_clinit, cl_context _context) {
+  context = _context;
+  if (do_clinit) clinit();
+  char *thneed_debug_env = getenv("THNEED_DEBUG");
+  debug = (thneed_debug_env != NULL) ? atoi(thneed_debug_env) : 0;
+}
+
+void Thneed::execute(float **finputs, float *foutput, bool slow) {
+  uint64_t tb, te;
+  if (debug >= 1) tb = nanos_since_boot();
+
+  // ****** copy inputs
+  copy_inputs(finputs);
+
+  // ****** run commands
+  clexec();
+
+  // ****** copy outputs
+  copy_output(foutput);
+
+  if (debug >= 1) {
+    te = nanos_since_boot();
+    printf("model exec in %lu us\n", (te-tb)/1000);
+  }
+}
--- a/sunnypilot/modeld/thneed/thneed_qcom2.cc
+++ b/sunnypilot/modeld/thneed/thneed_qcom2.cc
@@ -0,0 +1,258 @@
+#include "sunnypilot/modeld/thneed/thneed.h"
+
+#include <dlfcn.h>
+#include <sys/mman.h>
+
+#include <cassert>
+#include <cerrno>
+#include <cstring>
+#include <map>
+#include <string>
+
+#include "common/clutil.h"
+#include "common/timing.h"
+
+Thneed *g_thneed = NULL;
+int g_fd = -1;
+
+void hexdump(uint8_t *d, int len) {
+  assert((len%4) == 0);
+  printf("  dumping %p len 0x%x\n", d, len);
+  for (int i = 0; i < len/4; i++) {
+    if (i != 0 && (i%0x10) == 0) printf("\n");
+    printf("%8x ", d[i]);
+  }
+  printf("\n");
+}
+
+// *********** ioctl interceptor ***********
+
+extern "C" {
+
+int (*my_ioctl)(int filedes, unsigned long request, void *argp) = NULL;
+#undef ioctl
+int ioctl(int filedes, unsigned long request, void *argp) {
+  request &= 0xFFFFFFFF;  // needed on QCOM2
+  if (my_ioctl == NULL) my_ioctl = reinterpret_cast<decltype(my_ioctl)>(dlsym(RTLD_NEXT, "ioctl"));
+  Thneed *thneed = g_thneed;
+
+  // save the fd
+  if (request == IOCTL_KGSL_GPUOBJ_ALLOC) g_fd = filedes;
+
+  // note that this runs always, even without a thneed object
+  if (request == IOCTL_KGSL_DRAWCTXT_CREATE) {
+    struct kgsl_drawctxt_create *create = (struct kgsl_drawctxt_create *)argp;
+    create->flags &= ~KGSL_CONTEXT_PRIORITY_MASK;
+    create->flags |= 6 << KGSL_CONTEXT_PRIORITY_SHIFT;   // priority from 1-15, 1 is max priority
+    printf("IOCTL_KGSL_DRAWCTXT_CREATE: creating context with flags 0x%x\n", create->flags);
+  }
+
+  if (thneed != NULL) {
+    if (request == IOCTL_KGSL_GPU_COMMAND) {
+      struct kgsl_gpu_command *cmd = (struct kgsl_gpu_command *)argp;
+      if (thneed->record) {
+        thneed->timestamp = cmd->timestamp;
+        thneed->context_id = cmd->context_id;
+        thneed->cmds.push_back(unique_ptr<CachedCommand>(new CachedCommand(thneed, cmd)));
+      }
+      if (thneed->debug >= 1) {
+        printf("IOCTL_KGSL_GPU_COMMAND(%2zu): flags: 0x%lx    context_id: %u  timestamp: %u  numcmds: %d  numobjs: %d\n",
+            thneed->cmds.size(),
+            cmd->flags,
+            cmd->context_id, cmd->timestamp, cmd->numcmds, cmd->numobjs);
+      }
+    } else if (request == IOCTL_KGSL_GPUOBJ_SYNC) {
+      struct kgsl_gpuobj_sync *cmd = (struct kgsl_gpuobj_sync *)argp;
+      struct kgsl_gpuobj_sync_obj *objs = (struct kgsl_gpuobj_sync_obj *)(cmd->objs);
+
+      if (thneed->debug >= 2) {
+        printf("IOCTL_KGSL_GPUOBJ_SYNC count:%d ", cmd->count);
+        for (int i = 0; i < cmd->count; i++) {
+          printf(" -- offset:0x%lx len:0x%lx id:%d op:%d  ", objs[i].offset, objs[i].length, objs[i].id, objs[i].op);
+        }
+        printf("\n");
+      }
+
+      if (thneed->record) {
+        thneed->cmds.push_back(unique_ptr<CachedSync>(new
+              CachedSync(thneed, string((char *)objs, sizeof(struct kgsl_gpuobj_sync_obj)*cmd->count))));
+      }
+    } else if (request == IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID) {
+      struct kgsl_device_waittimestamp_ctxtid *cmd = (struct kgsl_device_waittimestamp_ctxtid *)argp;
+      if (thneed->debug >= 1) {
+        printf("IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID: context_id: %d  timestamp: %d  timeout: %d\n",
+            cmd->context_id, cmd->timestamp, cmd->timeout);
+      }
+    } else if (request == IOCTL_KGSL_SETPROPERTY) {
+      if (thneed->debug >= 1) {
+        struct kgsl_device_getproperty *prop = (struct kgsl_device_getproperty *)argp;
+        printf("IOCTL_KGSL_SETPROPERTY: 0x%x sizebytes:%zu\n", prop->type, prop->sizebytes);
+        if (thneed->debug >= 2) {
+          hexdump((uint8_t *)prop->value, prop->sizebytes);
+          if (prop->type == KGSL_PROP_PWR_CONSTRAINT) {
+            struct kgsl_device_constraint *constraint = (struct kgsl_device_constraint *)prop->value;
+            hexdump((uint8_t *)constraint->data, constraint->size);
+          }
+        }
+      }
+    } else if (request == IOCTL_KGSL_DRAWCTXT_CREATE || request == IOCTL_KGSL_DRAWCTXT_DESTROY) {
+      // this happens
+    } else if (request == IOCTL_KGSL_GPUOBJ_ALLOC || request == IOCTL_KGSL_GPUOBJ_FREE) {
+      // this happens
+    } else {
+      if (thneed->debug >= 1) {
+        printf("other ioctl %lx\n", request);
+      }
+    }
+  }
+
+  int ret = my_ioctl(filedes, request, argp);
+  // NOTE: This error message goes into stdout and messes up pyenv
+  // if (ret != 0) printf("ioctl returned %d with errno %d\n", ret, errno);
+  return ret;
+}
+
+}
+
+// *********** GPUMalloc ***********
+
+GPUMalloc::GPUMalloc(int size, int fd) {
+  struct kgsl_gpuobj_alloc alloc;
+  memset(&alloc, 0, sizeof(alloc));
+  alloc.size = size;
+  alloc.flags = 0x10000a00;
+  ioctl(fd, IOCTL_KGSL_GPUOBJ_ALLOC, &alloc);
+  void *addr = mmap64(NULL, alloc.mmapsize, 0x3, 0x1, fd, alloc.id*0x1000);
+  assert(addr != MAP_FAILED);
+
+  base = (uint64_t)addr;
+  remaining = size;
+}
+
+GPUMalloc::~GPUMalloc() {
+  // TODO: free the GPU malloced area
+}
+
+void *GPUMalloc::alloc(int size) {
+  void *ret = (void*)base;
+  size = (size+0xff) & (~0xFF);
+  assert(size <= remaining);
+  remaining -= size;
+  base += size;
+  return ret;
+}
+
+// *********** CachedSync, at the ioctl layer ***********
+
+void CachedSync::exec() {
+  struct kgsl_gpuobj_sync cmd;
+
+  cmd.objs = (uint64_t)data.data();
+  cmd.obj_len = data.length();
+  cmd.count = data.length() / sizeof(struct kgsl_gpuobj_sync_obj);
+
+  int ret = ioctl(thneed->fd, IOCTL_KGSL_GPUOBJ_SYNC, &cmd);
+  assert(ret == 0);
+}
+
+// *********** CachedCommand, at the ioctl layer ***********
+
+CachedCommand::CachedCommand(Thneed *lthneed, struct kgsl_gpu_command *cmd) {
+  thneed = lthneed;
+  assert(cmd->numsyncs == 0);
+
+  memcpy(&cache, cmd, sizeof(cache));
+
+  if (cmd->numcmds > 0) {
+    cmds = make_unique<struct kgsl_command_object[]>(cmd->numcmds);
+    memcpy(cmds.get(), (void *)cmd->cmdlist, sizeof(struct kgsl_command_object)*cmd->numcmds);
+    cache.cmdlist = (uint64_t)cmds.get();
+    for (int i = 0; i < cmd->numcmds; i++) {
+      void *nn = thneed->ram->alloc(cmds[i].size);
+      memcpy(nn, (void*)cmds[i].gpuaddr, cmds[i].size);
+      cmds[i].gpuaddr = (uint64_t)nn;
+    }
+  }
+
+  if (cmd->numobjs > 0) {
+    objs = make_unique<struct kgsl_command_object[]>(cmd->numobjs);
+    memcpy(objs.get(), (void *)cmd->objlist, sizeof(struct kgsl_command_object)*cmd->numobjs);
+    cache.objlist = (uint64_t)objs.get();
+    for (int i = 0; i < cmd->numobjs; i++) {
+      void *nn = thneed->ram->alloc(objs[i].size);
+      memset(nn, 0, objs[i].size);
+      objs[i].gpuaddr = (uint64_t)nn;
+    }
+  }
+
+  kq = thneed->ckq;
+  thneed->ckq.clear();
+}
+
+void CachedCommand::exec() {
+  cache.timestamp = ++thneed->timestamp;
+  int ret = ioctl(thneed->fd, IOCTL_KGSL_GPU_COMMAND, &cache);
+
+  if (thneed->debug >= 1) printf("CachedCommand::exec got %d\n", ret);
+
+  if (thneed->debug >= 2) {
+    for (auto &it : kq) {
+      it->debug_print(false);
+    }
+  }
+
+  assert(ret == 0);
+}
+
+// *********** Thneed ***********
+
+Thneed::Thneed(bool do_clinit, cl_context _context) {
+  // TODO: QCOM2 actually requires a different context
+  //context = _context;
+  if (do_clinit) clinit();
+  assert(g_fd != -1);
+  fd = g_fd;
+  ram = make_unique<GPUMalloc>(0x80000, fd);
+  timestamp = -1;
+  g_thneed = this;
+  char *thneed_debug_env = getenv("THNEED_DEBUG");
+  debug = (thneed_debug_env != NULL) ? atoi(thneed_debug_env) : 0;
+}
+
+void Thneed::wait() {
+  struct kgsl_device_waittimestamp_ctxtid wait;
+  wait.context_id = context_id;
+  wait.timestamp = timestamp;
+  wait.timeout = -1;
+
+  uint64_t tb = nanos_since_boot();
+  int wret = ioctl(fd, IOCTL_KGSL_DEVICE_WAITTIMESTAMP_CTXTID, &wait);
+  uint64_t te = nanos_since_boot();
+
+  if (debug >= 1) printf("wait %d after %lu us\n", wret, (te-tb)/1000);
+}
+
+void Thneed::execute(float **finputs, float *foutput, bool slow) {
+  uint64_t tb, te;
+  if (debug >= 1) tb = nanos_since_boot();
+
+  // ****** copy inputs
+  copy_inputs(finputs, true);
+
+  // ****** run commands
+  int i = 0;
+  for (auto &it : cmds) {
+    ++i;
+    if (debug >= 1) printf("run %2d @ %7lu us: ", i, (nanos_since_boot()-tb)/1000);
+    it->exec();
+    if ((i == cmds.size()) || slow) wait();
+  }
+
+  // ****** copy outputs
+  copy_output(foutput);
+
+  if (debug >= 1) {
+    te = nanos_since_boot();
+    printf("model exec in %lu us\n", (te-tb)/1000);
+  }
+}
--- a/sunnypilot/modeld/transforms/loadyuv.cc
+++ b/sunnypilot/modeld/transforms/loadyuv.cc
@@ -0,0 +1,76 @@
+#include "selfdrive/modeld/transforms/loadyuv.h"
+
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+
+void loadyuv_init(LoadYUVState* s, cl_context ctx, cl_device_id device_id, int width, int height) {
+  memset(s, 0, sizeof(*s));
+
+  s->width = width;
+  s->height = height;
+
+  char args[1024];
+  snprintf(args, sizeof(args),
+           "-cl-fast-relaxed-math -cl-denorms-are-zero "
+           "-DTRANSFORMED_WIDTH=%d -DTRANSFORMED_HEIGHT=%d",
+           width, height);
+  cl_program prg = cl_program_from_file(ctx, device_id, LOADYUV_PATH, args);
+
+  s->loadys_krnl = CL_CHECK_ERR(clCreateKernel(prg, "loadys", &err));
+  s->loaduv_krnl = CL_CHECK_ERR(clCreateKernel(prg, "loaduv", &err));
+  s->copy_krnl = CL_CHECK_ERR(clCreateKernel(prg, "copy", &err));
+
+  // done with this
+  CL_CHECK(clReleaseProgram(prg));
+}
+
+void loadyuv_destroy(LoadYUVState* s) {
+  CL_CHECK(clReleaseKernel(s->loadys_krnl));
+  CL_CHECK(clReleaseKernel(s->loaduv_krnl));
+  CL_CHECK(clReleaseKernel(s->copy_krnl));
+}
+
+void loadyuv_queue(LoadYUVState* s, cl_command_queue q,
+                   cl_mem y_cl, cl_mem u_cl, cl_mem v_cl,
+                   cl_mem out_cl) {
+  cl_int global_out_off = 0;
+
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 0, sizeof(cl_mem), &y_cl));
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loadys_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  const size_t loadys_work_size = (s->width*s->height)/8;
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loadys_krnl, 1, NULL,
+                               &loadys_work_size, NULL, 0, 0, NULL));
+
+  const size_t loaduv_work_size = ((s->width/2)*(s->height/2))/8;
+  global_out_off += (s->width*s->height);
+
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 0, sizeof(cl_mem), &u_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loaduv_krnl, 1, NULL,
+                               &loaduv_work_size, NULL, 0, 0, NULL));
+
+  global_out_off += (s->width/2)*(s->height/2);
+
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 0, sizeof(cl_mem), &v_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 1, sizeof(cl_mem), &out_cl));
+  CL_CHECK(clSetKernelArg(s->loaduv_krnl, 2, sizeof(cl_int), &global_out_off));
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->loaduv_krnl, 1, NULL,
+                               &loaduv_work_size, NULL, 0, 0, NULL));
+}
+
+void copy_queue(LoadYUVState* s, cl_command_queue q, cl_mem src, cl_mem dst,
+                 size_t src_offset, size_t dst_offset, size_t size) {
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 0, sizeof(cl_mem), &src));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 1, sizeof(cl_mem), &dst));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 2, sizeof(cl_int), &src_offset));
+  CL_CHECK(clSetKernelArg(s->copy_krnl, 3, sizeof(cl_int), &dst_offset));
+  const size_t copy_work_size = size/8;
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->copy_krnl, 1, NULL,
+                              &copy_work_size, NULL, 0, 0, NULL));
+}
--- a/sunnypilot/modeld/transforms/loadyuv.cl
+++ b/sunnypilot/modeld/transforms/loadyuv.cl
@@ -0,0 +1,47 @@
+#define UV_SIZE ((TRANSFORMED_WIDTH/2)*(TRANSFORMED_HEIGHT/2))
+
+__kernel void loadys(__global uchar8 const * const Y,
+                     __global uchar * out,
+                     int out_offset)
+{
+    const int gid = get_global_id(0);
+    const int ois = gid * 8;
+    const int oy = ois / TRANSFORMED_WIDTH;
+    const int ox = ois % TRANSFORMED_WIDTH;
+
+    const uchar8 ys = Y[gid];
+
+    // 02
+    // 13
+
+    __global uchar* outy0;
+    __global uchar* outy1;
+    if ((oy & 1) == 0) {
+      outy0 = out + out_offset; //y0
+      outy1 = out + out_offset + UV_SIZE*2; //y2
+    } else {
+      outy0 = out + out_offset + UV_SIZE; //y1
+      outy1 = out + out_offset + UV_SIZE*3; //y3
+    }
+
+    vstore4(ys.s0246, 0, outy0 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+    vstore4(ys.s1357, 0, outy1 + (oy/2) * (TRANSFORMED_WIDTH/2) + ox/2);
+}
+
+__kernel void loaduv(__global uchar8 const * const in,
+                     __global uchar8 * out,
+                     int out_offset)
+{
+  const int gid = get_global_id(0);
+  const uchar8 inv = in[gid];
+  out[gid + out_offset / 8] = inv;
+}
+
+__kernel void copy(__global uchar8 * in,
+                   __global uchar8 * out,
+                   int in_offset,
+                   int out_offset)
+{
+  const int gid = get_global_id(0);
+  out[gid + out_offset / 8] = in[gid + in_offset / 8];
+}
--- a/sunnypilot/modeld/transforms/loadyuv.h
+++ b/sunnypilot/modeld/transforms/loadyuv.h
@@ -0,0 +1,20 @@
+#pragma once
+
+#include "common/clutil.h"
+
+typedef struct {
+  int width, height;
+  cl_kernel loadys_krnl, loaduv_krnl, copy_krnl;
+} LoadYUVState;
+
+void loadyuv_init(LoadYUVState* s, cl_context ctx, cl_device_id device_id, int width, int height);
+
+void loadyuv_destroy(LoadYUVState* s);
+
+void loadyuv_queue(LoadYUVState* s, cl_command_queue q,
+                   cl_mem y_cl, cl_mem u_cl, cl_mem v_cl,
+                   cl_mem out_cl);
+
+
+void copy_queue(LoadYUVState* s, cl_command_queue q, cl_mem src, cl_mem dst,
+                 size_t src_offset, size_t dst_offset, size_t size);
--- a/sunnypilot/modeld/transforms/transform.cc
+++ b/sunnypilot/modeld/transforms/transform.cc
@@ -0,0 +1,97 @@
+#include "selfdrive/modeld/transforms/transform.h"
+
+#include <cassert>
+#include <cstring>
+
+#include "common/clutil.h"
+
+void transform_init(Transform* s, cl_context ctx, cl_device_id device_id) {
+  memset(s, 0, sizeof(*s));
+
+  cl_program prg = cl_program_from_file(ctx, device_id, TRANSFORM_PATH, "");
+  s->krnl = CL_CHECK_ERR(clCreateKernel(prg, "warpPerspective", &err));
+  // done with this
+  CL_CHECK(clReleaseProgram(prg));
+
+  s->m_y_cl = CL_CHECK_ERR(clCreateBuffer(ctx, CL_MEM_READ_WRITE, 3*3*sizeof(float), NULL, &err));
+  s->m_uv_cl = CL_CHECK_ERR(clCreateBuffer(ctx, CL_MEM_READ_WRITE, 3*3*sizeof(float), NULL, &err));
+}
+
+void transform_destroy(Transform* s) {
+  CL_CHECK(clReleaseMemObject(s->m_y_cl));
+  CL_CHECK(clReleaseMemObject(s->m_uv_cl));
+  CL_CHECK(clReleaseKernel(s->krnl));
+}
+
+void transform_queue(Transform* s,
+                     cl_command_queue q,
+                     cl_mem in_yuv, int in_width, int in_height, int in_stride, int in_uv_offset,
+                     cl_mem out_y, cl_mem out_u, cl_mem out_v,
+                     int out_width, int out_height,
+                     const mat3& projection) {
+  const int zero = 0;
+
+  // sampled using pixel center origin
+  // (because that's how fastcv and opencv does it)
+
+  mat3 projection_y = projection;
+
+  // in and out uv is half the size of y.
+  mat3 projection_uv = transform_scale_buffer(projection, 0.5);
+
+  CL_CHECK(clEnqueueWriteBuffer(q, s->m_y_cl, CL_TRUE, 0, 3*3*sizeof(float), (void*)projection_y.v, 0, NULL, NULL));
+  CL_CHECK(clEnqueueWriteBuffer(q, s->m_uv_cl, CL_TRUE, 0, 3*3*sizeof(float), (void*)projection_uv.v, 0, NULL, NULL));
+
+  const int in_y_width = in_width;
+  const int in_y_height = in_height;
+  const int in_y_px_stride = 1;
+  const int in_uv_width = in_width/2;
+  const int in_uv_height = in_height/2;
+  const int in_uv_px_stride = 2;
+  const int in_u_offset = in_uv_offset;
+  const int in_v_offset = in_uv_offset + 1;
+
+  const int out_y_width = out_width;
+  const int out_y_height = out_height;
+  const int out_uv_width = out_width/2;
+  const int out_uv_height = out_height/2;
+
+  CL_CHECK(clSetKernelArg(s->krnl, 0, sizeof(cl_mem), &in_yuv));  // src
+  CL_CHECK(clSetKernelArg(s->krnl, 1, sizeof(cl_int), &in_stride));  // src_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 2, sizeof(cl_int), &in_y_px_stride));  // src_px_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &zero));  // src_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 4, sizeof(cl_int), &in_y_height));  // src_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 5, sizeof(cl_int), &in_y_width));  // src_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_y));  // dst
+  CL_CHECK(clSetKernelArg(s->krnl, 7, sizeof(cl_int), &out_y_width));  // dst_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 8, sizeof(cl_int), &zero));  // dst_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 9, sizeof(cl_int), &out_y_height));  // dst_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 10, sizeof(cl_int), &out_y_width));  // dst_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 11, sizeof(cl_mem), &s->m_y_cl));  // M
+
+  const size_t work_size_y[2] = {(size_t)out_y_width, (size_t)out_y_height};
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_y, NULL, 0, 0, NULL));
+
+  const size_t work_size_uv[2] = {(size_t)out_uv_width, (size_t)out_uv_height};
+
+  CL_CHECK(clSetKernelArg(s->krnl, 2, sizeof(cl_int), &in_uv_px_stride));  // src_px_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &in_u_offset));  // src_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 4, sizeof(cl_int), &in_uv_height));  // src_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 5, sizeof(cl_int), &in_uv_width));  // src_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_u));  // dst
+  CL_CHECK(clSetKernelArg(s->krnl, 7, sizeof(cl_int), &out_uv_width));  // dst_row_stride
+  CL_CHECK(clSetKernelArg(s->krnl, 8, sizeof(cl_int), &zero));  // dst_offset
+  CL_CHECK(clSetKernelArg(s->krnl, 9, sizeof(cl_int), &out_uv_height));  // dst_rows
+  CL_CHECK(clSetKernelArg(s->krnl, 10, sizeof(cl_int), &out_uv_width));  // dst_cols
+  CL_CHECK(clSetKernelArg(s->krnl, 11, sizeof(cl_mem), &s->m_uv_cl));  // M
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_uv, NULL, 0, 0, NULL));
+  CL_CHECK(clSetKernelArg(s->krnl, 3, sizeof(cl_int), &in_v_offset));  // src_ofset
+  CL_CHECK(clSetKernelArg(s->krnl, 6, sizeof(cl_mem), &out_v));  // dst
+
+  CL_CHECK(clEnqueueNDRangeKernel(q, s->krnl, 2, NULL,
+                              (const size_t*)&work_size_uv, NULL, 0, 0, NULL));
+}
--- a/sunnypilot/modeld/transforms/transform.cl
+++ b/sunnypilot/modeld/transforms/transform.cl
@@ -0,0 +1,54 @@
+#define INTER_BITS 5
+#define INTER_TAB_SIZE (1 << INTER_BITS)
+#define INTER_SCALE 1.f / INTER_TAB_SIZE
+
+#define INTER_REMAP_COEF_BITS 15
+#define INTER_REMAP_COEF_SCALE (1 << INTER_REMAP_COEF_BITS)
+
+__kernel void warpPerspective(__global const uchar * src,
+                              int src_row_stride, int src_px_stride, int src_offset, int src_rows, int src_cols,
+                              __global uchar * dst,
+                              int dst_row_stride, int dst_offset, int dst_rows, int dst_cols,
+                              __constant float * M)
+{
+    int dx = get_global_id(0);
+    int dy = get_global_id(1);
+
+    if (dx < dst_cols && dy < dst_rows)
+    {
+        float X0 = M[0] * dx + M[1] * dy + M[2];
+        float Y0 = M[3] * dx + M[4] * dy + M[5];
+        float W = M[6] * dx + M[7] * dy + M[8];
+        W = W != 0.0f ? INTER_TAB_SIZE / W : 0.0f;
+        int X = rint(X0 * W), Y = rint(Y0 * W);
+
+        int sx = convert_short_sat(X >> INTER_BITS);
+        int sy = convert_short_sat(Y >> INTER_BITS);
+
+        short sx_clamp = clamp(sx, 0, src_cols - 1);
+        short sx_p1_clamp = clamp(sx + 1, 0, src_cols - 1);
+        short sy_clamp = clamp(sy, 0, src_rows - 1);
+        short sy_p1_clamp = clamp(sy + 1, 0, src_rows - 1);
+        int v0 = convert_int(src[mad24(sy_clamp, src_row_stride, src_offset + sx_clamp*src_px_stride)]);
+        int v1 = convert_int(src[mad24(sy_clamp, src_row_stride, src_offset + sx_p1_clamp*src_px_stride)]);
+        int v2 = convert_int(src[mad24(sy_p1_clamp, src_row_stride, src_offset + sx_clamp*src_px_stride)]);
+        int v3 = convert_int(src[mad24(sy_p1_clamp, src_row_stride, src_offset + sx_p1_clamp*src_px_stride)]);
+
+        short ay = (short)(Y & (INTER_TAB_SIZE - 1));
+        short ax = (short)(X & (INTER_TAB_SIZE - 1));
+        float taby = 1.f/INTER_TAB_SIZE*ay;
+        float tabx = 1.f/INTER_TAB_SIZE*ax;
+
+        int dst_index = mad24(dy, dst_row_stride, dst_offset + dx);
+
+        int itab0 = convert_short_sat_rte( (1.0f-taby)*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab1 = convert_short_sat_rte( (1.0f-taby)*tabx * INTER_REMAP_COEF_SCALE );
+        int itab2 = convert_short_sat_rte( taby*(1.0f-tabx) * INTER_REMAP_COEF_SCALE );
+        int itab3 = convert_short_sat_rte( taby*tabx * INTER_REMAP_COEF_SCALE );
+
+        int val = v0 * itab0 +  v1 * itab1 + v2 * itab2 + v3 * itab3;
+
+        uchar pix = convert_uchar_sat((val + (1 << (INTER_REMAP_COEF_BITS-1))) >> INTER_REMAP_COEF_BITS);
+        dst[dst_index] = pix;
+    }
+}
--- a/sunnypilot/modeld/transforms/transform.h
+++ b/sunnypilot/modeld/transforms/transform.h
@@ -0,0 +1,25 @@
+#pragma once
+
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#ifdef __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+#include "common/mat.h"
+
+typedef struct {
+  cl_kernel krnl;
+  cl_mem m_y_cl, m_uv_cl;
+} Transform;
+
+void transform_init(Transform* s, cl_context ctx, cl_device_id device_id);
+
+void transform_destroy(Transform* transform);
+
+void transform_queue(Transform* s, cl_command_queue q,
+                     cl_mem yuv, int in_width, int in_height, int in_stride, int in_uv_offset,
+                     cl_mem out_y, cl_mem out_u, cl_mem out_v,
+                     int out_width, int out_height,
+                     const mat3& projection);
--- a/sunnypilot/models/helpers.py
+++ b/sunnypilot/models/helpers.py
@@ -21,6 +21,7 @@ async def verify_file(file_path: str, expected_hash: str) -> bool:

  return sha256_hash.hexdigest().lower() == expected_hash.lower()

+
 def get_active_bundle(params: Params) -> custom.ModelManagerSP.ModelBundle:
  """Gets the active model bundle from cache"""
  if params is None:
@@ -30,3 +31,23 @@ def get_active_bundle(params: Params) -> custom.ModelManagerSP.ModelBundle:
    return messaging.log_from_bytes(active_bundle, custom.ModelManagerSP.ModelBundle)

  return None
+
+
+def get_model_runner_by_filename(filename: str) -> custom.ModelManagerSP.Runner:
+  if filename.endswith(".thneed"):
+    return custom.ModelManagerSP.Runner.snpe
+
+  if filename.endswith("_tinygrad.pkl"):
+    return custom.ModelManagerSP.Runner.tinygrad
+
+
+def get_active_model_runner(params: Params) -> custom.ModelManagerSP.Runner:
+  """Gets the model runner from the active model bundle. If no active bundle, returns tinygrad"""
+  if params is None:
+    params = Params()
+
+  if active_bundle := get_active_bundle(params):
+    drive_model = next(model for model in active_bundle.models if model.type == custom.ModelManagerSP.Type.drive)
+    return get_model_runner_by_filename(drive_model.fileName)
+
+  return custom.ModelManagerSP.Runner.tinygrad
--- a/system/manager/process_config.py
+++ b/system/manager/process_config.py
@@ -1,10 +1,12 @@
 import os
 import operator

-from cereal import car
+from cereal import car, custom
 from openpilot.common.params import Params
 from openpilot.system.hardware import PC, TICI
 from openpilot.system.manager.process import PythonProcess, NativeProcess, DaemonProcess
+
+from sunnypilot.models.helpers import get_active_model_runner
 from sunnypilot.sunnylink.utils import sunnylink_need_register, sunnylink_ready, use_sunnylink_uploader

 WEBCAM = os.getenv("USE_WEBCAM") is not None
@@ -70,6 +72,15 @@ def use_sunnylink_uploader_shim(started, params, CP: car.CarParams) -> bool:
  """Shim for use_sunnylink_uploader to match the process manager signature."""
  return use_sunnylink_uploader(params)

+def is_snpe_model(started, params, CP: car.CarParams) -> bool:
+  """Check if the active model runner is SNPE."""
+  # TODO-SP: I want to do a little more optimization here to only check this once when we've transitioned from offroad to onroad.
+  return bool(get_active_model_runner(params) == custom.ModelManagerSP.Runner.snpe)
+
+def is_stock_model(started, params, CP: car.CarParams) -> bool:
+  """Check if the active model runner is stock."""
+  return not is_snpe_model(started, params, CP)
+
 def or_(*fns):
  return lambda *args: operator.or_(*(fn(*args) for fn in fns))

@@ -92,7 +103,7 @@ procs = [
  NativeProcess("stream_encoderd", "system/loggerd", ["./encoderd", "--stream"], notcar),
  NativeProcess("loggerd", "system/loggerd", ["./loggerd"], logging),
  # TODO Make python process once TG allows opening QCOM from child proc
-  NativeProcess("modeld", "selfdrive/modeld", ["./modeld"], only_onroad),
+  NativeProcess("modeld", "selfdrive/modeld", ["./modeld"], and_(only_onroad, is_stock_model)),
  NativeProcess("sensord", "system/sensord", ["./sensord"], only_onroad, enabled=not PC),
  NativeProcess("ui", "selfdrive/ui", ["./ui"], always_run, watchdog_max_dt=(5 if not PC else None)),
  PythonProcess("soundd", "selfdrive.ui.soundd", only_onroad),
@@ -134,6 +145,7 @@ procs = [
 # sunnypilot
 procs += [
  PythonProcess("models_manager", "sunnypilot.models.manager", only_offroad),
+  NativeProcess("modeld_snpe", "sunnypilot/modeld", ["./modeld"], and_(only_onroad, is_snpe_model)),
 ]

 if os.path.exists("./github_runner.sh"):