Merge branch 'master' into tinygrad-sync-4/25

This reverts commit dc11e5fd84.

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

@@ -172,7 +172,7 @@ jobs:
             output_file="${{ env.MODELS_DIR }}/${base_name}_tinygrad.pkl"
 
             echo "Compiling: $onnx_file -> $output_file"
-            QCOM=1 python3 "${{ env.TINYGRAD_PATH }}/examples/openpilot/compile3.py" "$onnx_file" "$output_file"
+            DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 IMAGE=2 python3 "${{ env.TINYGRAD_PATH }}/examples/openpilot/compile3.py" "$onnx_file" "$output_file"
             DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 python3 "${{ env.MODELS_DIR }}/../get_model_metadata.py" "$onnx_file" || true
           done
 

selfdrive/modeld/SConscript

@@ -1,10 +1,23 @@
 import glob
 import json
 import os
+from itertools import product
 from SCons.Script import Value
 from openpilot.common.file_chunker import chunk_file, get_chunk_paths
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE, DM_INPUT_SIZE
+from openpilot.selfdrive.modeld.constants import ModelConstants
+from openpilot.selfdrive.modeld.helpers import CompileConfig
 from tinygrad import Device
 
+CAMERA_CONFIGS = [
+  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
+  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
+]
+MODELD_CONFIGS = [CompileConfig(cam_w, cam_h, prepare_only, 'driving_')
+                  for (cam_w, cam_h), prepare_only in product(CAMERA_CONFIGS, [True, False])]
+DM_WARP_CONFIGS = [CompileConfig(cam_w, cam_h, True, 'dm_') for cam_w, cam_h in CAMERA_CONFIGS]
+
 Import('env', 'arch')
 chunker_file = File("#common/file_chunker.py")
 lenv = env.Clone()
@@ -16,18 +29,17 @@ tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "
 def estimate_pickle_max_size(onnx_size):
   return 1.2 * onnx_size  + 10 * 1024 * 1024  # 20% + 10MB is plenty
 
-# THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
 # get fastest TG config
 available = set(Device.get_available_devices())
-# FIXME-SP: reset when we bump tg
+if 'CUDA' in available:
-if False:  # 'CUDA' in available:
   tg_backend = 'CUDA'
   tg_flags = f'DEV={tg_backend}'
 elif 'QCOM' in available:
   tg_backend = 'QCOM'
-  tg_flags = f'DEV={tg_backend} FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0'
+  tg_flags = f'DEV={tg_backend} FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
 else:
-  tg_backend = 'CPU' if arch == 'Darwin' else 'CPU CPU_LLVM=1'  # FIXME-SP: reset when we bump tg
+  tg_backend = 'CPU' if arch == 'Darwin' else 'CPU:LLVM'
+  # THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
   tg_flags = f'DEV={tg_backend} THREADS=0'
 
 def write_tg_compiled_flags(target, source, env):
@@ -54,14 +66,35 @@ for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
 image_flag = {
      'larch64': 'IMAGE=2',
 }.get(arch, 'IMAGE=0')
-script_files = [File(Dir("#selfdrive/modeld").File("compile_warp.py").abspath)]
+modeld_dir = Dir("#selfdrive/modeld").abspath
-compile_warp_cmd = f'{tg_flags} {mac_brew_string} python3 {Dir("#selfdrive/modeld").abspath}/compile_warp.py '
+compile_modeld_script = [File(f"{modeld_dir}/compile_modeld.py")]
-from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
+compile_dm_warp_script = [File(f"{modeld_dir}/compile_dm_warp.py")]
-warp_targets = []
+driving_onnx_deps = [File(f"models/{m}.onnx").abspath for m in ['driving_vision', 'driving_policy']]
-for cam in [_ar_ox_fisheye, _os_fisheye]:
+driving_metadata_deps = [File(f"models/{m}_metadata.pkl").abspath for m in ['driving_vision', 'driving_policy']]
-  w, h = cam.width, cam.height
+
-  warp_targets += [File(f"models/warp_{w}x{h}_tinygrad.pkl").abspath, File(f"models/dm_warp_{w}x{h}_tinygrad.pkl").abspath]
+model_w, model_h = MEDMODEL_INPUT_SIZE
-lenv.Command(warp_targets, tinygrad_files + script_files + [compiled_flags_node], compile_warp_cmd)
+frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
+for cfg in MODELD_CONFIGS:
+  cmd = (f'{tg_flags} {mac_brew_string} {image_flag} python3 {modeld_dir}/compile_modeld.py '
+         f'--model-size {model_w}x{model_h} '
+         f'--nv12 {",".join(str(x) for x in cfg.nv12)} '
+         f'--vision-onnx {File("models/driving_vision.onnx").abspath} '
+         f'--policy-onnx {File("models/driving_policy.onnx").abspath} '
+         f'--output {cfg.pkl_path} --frame-skip {frame_skip}'
+         + (' --prepare-only' if cfg.prepare_only else ''))
+  node = lenv.Command(cfg.pkl_path, tinygrad_files + compile_modeld_script + driving_onnx_deps + driving_metadata_deps + [chunker_file, compiled_flags_node], cmd)
+  onnx_sizes_sum = sum(os.path.getsize(f) for f in driving_onnx_deps)
+  chunk_targets = get_chunk_paths(cfg.pkl_path, estimate_pickle_max_size(onnx_sizes_sum))
+  def do_chunk(target, source, env, pkl=cfg.pkl_path, chunks=chunk_targets):
+    chunk_file(pkl, chunks)
+  lenv.Command(chunk_targets, node, do_chunk)
+
+dm_w, dm_h = DM_INPUT_SIZE
+for cfg in DM_WARP_CONFIGS:
+  cmd = (f'{tg_flags} {mac_brew_string} {image_flag} python3 {modeld_dir}/compile_dm_warp.py '
+         f'--nv12 {",".join(str(x) for x in cfg.nv12)} --warp-to {dm_w}x{dm_h} '
+         f'--output {cfg.pkl_path}')
+  lenv.Command(cfg.pkl_path, tinygrad_files + compile_dm_warp_script + compile_modeld_script + [compiled_flags_node], cmd)
 
 def tg_compile(flags, model_name):
   pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"'
@@ -82,7 +115,4 @@ def tg_compile(flags, model_name):
     do_chunk,
   )
 
-# Compile small models
+tg_compile(tg_flags, 'dmonitoring_model')
-for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
-  tg_compile(tg_flags, model_name)
-

selfdrive/modeld/compile_dm_warp.py

@@ -0,0 +1,54 @@
+#!/usr/bin/env python3
+import argparse
+import pickle
+import time
+
+from tinygrad.tensor import Tensor
+from tinygrad.device import Device
+from tinygrad.engine.jit import TinyJit
+
+from openpilot.selfdrive.modeld.compile_modeld import NV12Frame, warp_perspective_tinygrad, _parse_size, _parse_nv12
+
+
+def make_warp_dm(nv12: NV12Frame, dm_w, dm_h):
+  cam_w, cam_h, stride, _, _, _ = nv12
+  stride_pad = stride - cam_w
+
+  def warp_dm(input_frame, M_inv):
+    M_inv = M_inv.to(Device.DEFAULT).realize()
+    return warp_perspective_tinygrad(input_frame[:cam_h*stride], M_inv,
+                                     (dm_w, dm_h), (cam_h, cam_w), stride_pad).reshape(-1, dm_h * dm_w)
+  return warp_dm
+
+
+def compile_dm_warp(nv12: NV12Frame, dm_w, dm_h, pkl_path):
+  print(f"Compiling DM warp for {nv12.width}x{nv12.height} -> {dm_w}x{dm_h}...")
+
+  warp_dm_jit = TinyJit(make_warp_dm(nv12, dm_w, dm_h), prune=True)
+
+  for i in range(10):
+    frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
+    M_inv = Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')
+    Device.default.synchronize()
+    st = time.perf_counter()
+    warp_dm_jit(frame, M_inv).realize()
+    mt = time.perf_counter()
+    Device.default.synchronize()
+    et = time.perf_counter()
+    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
+
+  with open(pkl_path, "wb") as f:
+    pickle.dump(warp_dm_jit, f)
+  print(f"  Saved to {pkl_path}")
+
+
+if __name__ == "__main__":
+  p = argparse.ArgumentParser()
+  p.add_argument('--nv12', type=_parse_nv12, required=True,
+                 help=f'NV12 frame layout: {",".join(NV12Frame._fields)}')
+  p.add_argument('--warp-to', type=_parse_size, required=True, help='DM input WxH')
+  p.add_argument('--output', required=True)
+  args = p.parse_args()
+
+  dm_w, dm_h = args.warp_to
+  compile_dm_warp(args.nv12, dm_w, dm_h, args.output)

selfdrive/modeld/compile_modeld.py

@@ -0,0 +1,253 @@
+#!/usr/bin/env python3
+import argparse
+import pickle
+import time
+from functools import partial
+from collections import namedtuple
+
+import numpy as np
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import Context
+from tinygrad.device import Device
+from tinygrad.engine.jit import TinyJit
+from tinygrad.nn.onnx import OnnxRunner
+
+# https://github.com/tinygrad/tinygrad/issues/15682
+from tinygrad.uop.ops import UOp, Ops
+_orig = UOp.__reduce__
+UOp.__reduce__ = lambda self: (UOp.unique, ()) if self.op is Ops.UNIQUE else _orig(self)
+
+
+NV12Frame = namedtuple("NV12Frame", ['width', 'height', 'stride', 'y_height', 'uv_height', 'size'])
+
+UV_SCALE_MATRIX = np.array([[0.5, 0, 0], [0, 0.5, 0], [0, 0, 1]], dtype=np.float32)
+UV_SCALE_MATRIX_INV = np.linalg.inv(UV_SCALE_MATRIX)
+
+
+def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad):
+  w_dst, h_dst = dst_shape
+  h_src, w_src = src_shape
+
+  x = Tensor.arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
+  y = Tensor.arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst).reshape(-1)
+
+  # inline 3x3 matmul as elementwise to avoid reduce op (enables fusion with gather)
+  src_x = M_inv[0, 0] * x + M_inv[0, 1] * y + M_inv[0, 2]
+  src_y = M_inv[1, 0] * x + M_inv[1, 1] * y + M_inv[1, 2]
+  src_w = M_inv[2, 0] * x + M_inv[2, 1] * y + M_inv[2, 2]
+
+  src_x = src_x / src_w
+  src_y = src_y / src_w
+
+  x_nn_clipped = Tensor.round(src_x).clip(0, w_src - 1).cast('int')
+  y_nn_clipped = Tensor.round(src_y).clip(0, h_src - 1).cast('int')
+  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
+
+  return src_flat[idx]
+
+
+def frames_to_tensor(frames):
+  H = (frames.shape[0] * 2) // 3
+  W = frames.shape[1]
+  in_img1 = Tensor.cat(frames[0:H:2, 0::2],
+                       frames[1:H:2, 0::2],
+                       frames[0:H:2, 1::2],
+                       frames[1:H:2, 1::2],
+                       frames[H:H+H//4].reshape((H//2, W//2)),
+                       frames[H+H//4:H+H//2].reshape((H//2, W//2)), dim=0).reshape((6, H//2, W//2))
+  return in_img1
+
+
+def make_frame_prepare(nv12: NV12Frame, model_w, model_h):
+  cam_w, cam_h, stride, y_height, uv_height, _ = nv12
+  uv_offset = stride * y_height
+  stride_pad = stride - cam_w
+
+  def frame_prepare_tinygrad(input_frame, M_inv):
+    # UV_SCALE @ M_inv @ UV_SCALE_INV simplifies to elementwise scaling
+    M_inv_uv = M_inv * Tensor([[1.0, 1.0, 0.5], [1.0, 1.0, 0.5], [2.0, 2.0, 1.0]])
+    # deinterleave NV12 UV plane (UVUV... -> separate U, V)
+    uv = input_frame[uv_offset:uv_offset + uv_height * stride].reshape(uv_height, stride)
+    with Context(SPLIT_REDUCEOP=0):
+      y = warp_perspective_tinygrad(input_frame[:cam_h*stride],
+                                    M_inv, (model_w, model_h),
+                                    (cam_h, cam_w), stride_pad).realize()
+      u = warp_perspective_tinygrad(uv[:cam_h//2, :cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+      v = warp_perspective_tinygrad(uv[:cam_h//2, 1:cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+    yuv = y.cat(u).cat(v).reshape((model_h * 3 // 2, model_w))
+    tensor = frames_to_tensor(yuv)
+    return tensor
+  return frame_prepare_tinygrad
+
+
+def make_input_queues(vision_input_shapes, policy_input_shapes, frame_skip):
+  img = vision_input_shapes['img']  # (1, 12, 128, 256)
+  n_frames = img[1] // 6
+  img_buf_shape = (frame_skip * (n_frames - 1) + 1, 6, img[2], img[3])
+
+  fb = policy_input_shapes['features_buffer']  # (1, 25, 512)
+  dp = policy_input_shapes['desire_pulse']  # (1, 25, 8)
+  tc = policy_input_shapes['traffic_convention']  # (1, 2)
+
+  npy = {
+    'desire': np.zeros(dp[2], dtype=np.float32),
+    'traffic_convention': np.zeros(tc, dtype=np.float32),
+    'tfm': np.zeros((3, 3), dtype=np.float32),
+    'big_tfm': np.zeros((3, 3), dtype=np.float32),
+  }
+  input_queues = {
+    'img_q': Tensor.zeros(img_buf_shape, dtype='uint8').contiguous().realize(),
+    'big_img_q': Tensor.zeros(img_buf_shape, dtype='uint8').contiguous().realize(),
+    'feat_q': Tensor.zeros(frame_skip * (fb[1] - 1) + 1, fb[0], fb[2]).contiguous().realize(),
+    'desire_q': Tensor.zeros(frame_skip * dp[1], dp[0], dp[2]).contiguous().realize(),
+    **{k: Tensor(v, device='NPY').realize() for k, v in npy.items()},
+  }
+  return input_queues, npy
+
+
+def shift_and_sample(buf, new_val, sample_fn):
+  buf.assign(buf[1:].cat(new_val, dim=0).contiguous())
+  return sample_fn(buf)
+
+
+def sample_skip(buf, frame_skip):
+  return buf[::frame_skip].contiguous().flatten(0, 1).unsqueeze(0)
+
+
+def sample_desire(buf, frame_skip):
+  return buf.reshape(-1, frame_skip, *buf.shape[1:]).max(1).flatten(0, 1).unsqueeze(0)
+
+
+def make_run_policy(vision_runner, policy_runner, nv12: NV12Frame, model_w, model_h,
+                    vision_features_slice, frame_skip, prepare_only=False):
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
+  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
+  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
+
+  def run_policy(img_q, big_img_q, feat_q, desire_q, desire, traffic_convention, tfm, big_tfm, frame, big_frame):
+    tfm = tfm.to(Device.DEFAULT)
+    big_tfm = big_tfm.to(Device.DEFAULT)
+    desire = desire.to(Device.DEFAULT)
+    traffic_convention = traffic_convention.to(Device.DEFAULT)
+    Tensor.realize(tfm, big_tfm, desire, traffic_convention)
+
+    img = shift_and_sample(img_q, frame_prepare(frame, tfm).unsqueeze(0), sample_skip_fn)
+    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm).unsqueeze(0), sample_skip_fn)
+
+    if prepare_only:
+      return img, big_img
+
+    vision_out = next(iter(vision_runner({'img': img, 'big_img': big_img}).values())).cast('float32')
+
+    new_feat = vision_out[:, vision_features_slice].reshape(1, -1).unsqueeze(0)
+    feat_buf = shift_and_sample(feat_q, new_feat, sample_skip_fn)
+    desire_buf = shift_and_sample(desire_q, desire.reshape(1, 1, -1), sample_desire_fn)
+
+    inputs = {'features_buffer': feat_buf, 'desire_pulse': desire_buf, 'traffic_convention': traffic_convention}
+    policy_out = next(iter(policy_runner(inputs).values())).cast('float32')
+
+    return vision_out, policy_out
+  return run_policy
+
+
+def compile_modeld(nv12: NV12Frame, model_w, model_h, prepare_only, frame_skip,
+                   vision_onnx, policy_onnx, pkl_path):
+  from get_model_metadata import metadata_path_for
+
+  print(f"Compiling combined policy JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
+
+  vision_runner = OnnxRunner(vision_onnx)
+  policy_runner = OnnxRunner(policy_onnx)
+
+  with open(metadata_path_for(vision_onnx), 'rb') as f:
+    vision_metadata = pickle.load(f)
+    vision_features_slice = vision_metadata['output_slices']['hidden_state']
+    vision_input_shapes = vision_metadata['input_shapes']
+  with open(metadata_path_for(policy_onnx), 'rb') as f:
+    policy_input_shapes = pickle.load(f)['input_shapes']
+
+  _run = make_run_policy(vision_runner, policy_runner, nv12, model_w, model_h,
+                         vision_features_slice, frame_skip, prepare_only)
+  run_policy_jit = TinyJit(_run, prune=True)
+
+  N_RUNS = 3
+  SEED = 42
+
+  def random_inputs_run_fn(fn, seed, test_val=None, test_buffers=None, expect_match=True):
+    input_queues, npy = make_input_queues(vision_input_shapes, policy_input_shapes, frame_skip)
+    np.random.seed(seed)
+    Tensor.manual_seed(seed)
+
+    for i in range(N_RUNS):
+      frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
+      big_frame = Tensor.randint(nv12.size, low=0, high=256, dtype='uint8').realize()
+      for v in npy.values():
+        v[:] = np.random.randn(*v.shape).astype(v.dtype)
+      Device.default.synchronize()
+      st = time.perf_counter()
+      outs = fn(**input_queues, frame=frame, big_frame=big_frame)
+      mt = time.perf_counter()
+      for o in outs:
+        # .realize() not needed once jitted, but needed for unjitted fn
+        o.realize()
+      Device.default.synchronize()
+      et = time.perf_counter()
+      print(f"  [{i+1}/{N_RUNS}] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
+
+    val = [np.copy(v.numpy()) for v in outs]
+    buffers = [np.copy(v.numpy().copy()) for v in input_queues.values()]
+
+    if test_val is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(val, test_val, strict=True))
+      assert match == expect_match, f"outputs {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    if test_buffers is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(buffers, test_buffers, strict=True))
+      assert match == expect_match, f"buffers {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    return fn, val, buffers
+
+  print('run unjitted')
+  _, test_val, test_buffers = random_inputs_run_fn(_run, seed=SEED)
+  print('capture + replay')
+  run_policy_jit, _, _ = random_inputs_run_fn(run_policy_jit, SEED, test_val, test_buffers)
+
+  print('pickle round trip')
+  with open(pkl_path, "wb") as f:
+    pickle.dump(run_policy_jit, f)
+    print(f"  Saved to {pkl_path}")
+  with open(pkl_path, "rb") as f:
+    run_policy_jit = pickle.load(f)
+  random_inputs_run_fn(run_policy_jit, SEED, test_val, test_buffers, expect_match=True)
+  random_inputs_run_fn(run_policy_jit, SEED+1, test_val, test_buffers, expect_match=False)
+
+
+def _parse_size(s):
+  w, h = s.lower().split('x')
+  return int(w), int(h)
+
+
+def _parse_nv12(s):
+  parts = s.split(',')
+  assert len(parts) == len(NV12Frame._fields), \
+    f"--nv12 expects {','.join(NV12Frame._fields)} (got {s!r})"
+  return NV12Frame(*(int(x) for x in parts))
+
+
+if __name__ == "__main__":
+  p = argparse.ArgumentParser()
+  p.add_argument('--model-size', type=_parse_size, required=True, help='model input WxH')
+  p.add_argument('--nv12', type=_parse_nv12, required=True,
+                 help=f'NV12 frame layout: {",".join(NV12Frame._fields)}')
+  p.add_argument('--vision-onnx', required=True)
+  p.add_argument('--policy-onnx', required=True)
+  p.add_argument('--output', required=True)
+  p.add_argument('--prepare-only', action='store_true')
+  p.add_argument('--frame-skip', type=int, required=True)
+  args = p.parse_args()
+
+  model_w, model_h = args.model_size
+  compile_modeld(args.nv12, model_w, model_h, args.prepare_only, args.frame_skip,
+                 args.vision_onnx, args.policy_onnx, args.output)

selfdrive/modeld/compile_warp.py

@@ -1,201 +0,0 @@
-#!/usr/bin/env python3
-import time
-import pickle
-import numpy as np
-from pathlib import Path
-from tinygrad.tensor import Tensor
-from tinygrad.helpers import Context
-from tinygrad.device import Device
-from tinygrad.engine.jit import TinyJit
-
-from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE, DM_INPUT_SIZE
-from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
-
-MODELS_DIR = Path(__file__).parent / 'models'
-
-CAMERA_CONFIGS = [
-  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
-  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
-]
-
-UV_SCALE_MATRIX = np.array([[0.5, 0, 0], [0, 0.5, 0], [0, 0, 1]], dtype=np.float32)
-UV_SCALE_MATRIX_INV = np.linalg.inv(UV_SCALE_MATRIX)
-
-IMG_BUFFER_SHAPE = (30, MEDMODEL_INPUT_SIZE[1] // 2, MEDMODEL_INPUT_SIZE[0] // 2)
-
-
-def warp_pkl_path(w, h):
-  return MODELS_DIR / f'warp_{w}x{h}_tinygrad.pkl'
-
-
-def dm_warp_pkl_path(w, h):
-  return MODELS_DIR / f'dm_warp_{w}x{h}_tinygrad.pkl'
-
-
-def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad):
-  w_dst, h_dst = dst_shape
-  h_src, w_src = src_shape
-
-  x = Tensor.arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
-  y = Tensor.arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst).reshape(-1)
-
-  # inline 3x3 matmul as elementwise to avoid reduce op (enables fusion with gather)
-  src_x = M_inv[0, 0] * x + M_inv[0, 1] * y + M_inv[0, 2]
-  src_y = M_inv[1, 0] * x + M_inv[1, 1] * y + M_inv[1, 2]
-  src_w = M_inv[2, 0] * x + M_inv[2, 1] * y + M_inv[2, 2]
-
-  src_x = src_x / src_w
-  src_y = src_y / src_w
-
-  x_nn_clipped = Tensor.round(src_x).clip(0, w_src - 1).cast('int')
-  y_nn_clipped = Tensor.round(src_y).clip(0, h_src - 1).cast('int')
-  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
-
-  return src_flat[idx]
-
-
-def frames_to_tensor(frames, model_w, model_h):
-  H = (frames.shape[0] * 2) // 3
-  W = frames.shape[1]
-  in_img1 = Tensor.cat(frames[0:H:2, 0::2],
-                       frames[1:H:2, 0::2],
-                       frames[0:H:2, 1::2],
-                       frames[1:H:2, 1::2],
-                       frames[H:H+H//4].reshape((H//2, W//2)),
-                       frames[H+H//4:H+H//2].reshape((H//2, W//2)), dim=0).reshape((6, H//2, W//2))
-  return in_img1
-
-
-def make_frame_prepare(cam_w, cam_h, model_w, model_h):
-  stride, y_height, uv_height, _ = get_nv12_info(cam_w, cam_h)
-  uv_offset = stride * y_height
-  stride_pad = stride - cam_w
-
-  def frame_prepare_tinygrad(input_frame, M_inv):
-    # UV_SCALE @ M_inv @ UV_SCALE_INV simplifies to elementwise scaling
-    M_inv_uv = M_inv * Tensor([[1.0, 1.0, 0.5], [1.0, 1.0, 0.5], [2.0, 2.0, 1.0]])
-    # deinterleave NV12 UV plane (UVUV... -> separate U, V)
-    uv = input_frame[uv_offset:uv_offset + uv_height * stride].reshape(uv_height, stride)
-    with Context(SPLIT_REDUCEOP=0):
-      y = warp_perspective_tinygrad(input_frame[:cam_h*stride],
-                                    M_inv, (model_w, model_h),
-                                    (cam_h, cam_w), stride_pad).realize()
-      u = warp_perspective_tinygrad(uv[:cam_h//2, :cam_w:2].flatten(),
-                                    M_inv_uv, (model_w//2, model_h//2),
-                                    (cam_h//2, cam_w//2), 0).realize()
-      v = warp_perspective_tinygrad(uv[:cam_h//2, 1:cam_w:2].flatten(),
-                                    M_inv_uv, (model_w//2, model_h//2),
-                                    (cam_h//2, cam_w//2), 0).realize()
-    yuv = y.cat(u).cat(v).reshape((model_h * 3 // 2, model_w))
-    tensor = frames_to_tensor(yuv, model_w, model_h)
-    return tensor
-  return frame_prepare_tinygrad
-
-
-def make_update_img_input(frame_prepare, model_w, model_h):
-  def update_img_input_tinygrad(tensor, frame, M_inv):
-    M_inv = M_inv.to(Device.DEFAULT)
-    new_img = frame_prepare(frame, M_inv)
-    tensor.assign(tensor[6:].cat(new_img, dim=0).contiguous())
-    return Tensor.cat(tensor[:6], tensor[-6:], dim=0).contiguous().reshape(1, 12, model_h//2, model_w//2)
-  return update_img_input_tinygrad
-
-
-def make_update_both_imgs(frame_prepare, model_w, model_h):
-  update_img = make_update_img_input(frame_prepare, model_w, model_h)
-
-  def update_both_imgs_tinygrad(calib_img_buffer, new_img, M_inv,
-                                calib_big_img_buffer, new_big_img, M_inv_big):
-    calib_img_pair = update_img(calib_img_buffer, new_img, M_inv)
-    calib_big_img_pair = update_img(calib_big_img_buffer, new_big_img, M_inv_big)
-    return calib_img_pair, calib_big_img_pair
-  return update_both_imgs_tinygrad
-
-
-def make_warp_dm(cam_w, cam_h, dm_w, dm_h):
-  stride, y_height, _, _ = get_nv12_info(cam_w, cam_h)
-  stride_pad = stride - cam_w
-
-  def warp_dm(input_frame, M_inv):
-    M_inv = M_inv.to(Device.DEFAULT)
-    result = warp_perspective_tinygrad(input_frame[:cam_h*stride], M_inv, (dm_w, dm_h), (cam_h, cam_w), stride_pad).reshape(-1, dm_h * dm_w)
-    return result
-  return warp_dm
-
-
-def compile_modeld_warp(cam_w, cam_h):
-  model_w, model_h = MEDMODEL_INPUT_SIZE
-  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
-
-  print(f"Compiling modeld warp for {cam_w}x{cam_h}...")
-
-  frame_prepare = make_frame_prepare(cam_w, cam_h, model_w, model_h)
-  update_both_imgs = make_update_both_imgs(frame_prepare, model_w, model_h)
-  update_img_jit = TinyJit(update_both_imgs, prune=True)
-
-  full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
-  big_full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
-  new_frame_np = np.random.randint(0, 256, yuv_size, dtype=np.uint8)
-  new_big_frame_np = np.random.randint(0, 256, yuv_size, dtype=np.uint8)
-  for i in range(10):
-    img_inputs = [full_buffer,
-                  Tensor.from_blob(new_frame_np.ctypes.data, (yuv_size,), dtype='uint8').realize(),
-                  Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
-    big_img_inputs = [big_full_buffer,
-                      Tensor.from_blob(new_big_frame_np.ctypes.data, (yuv_size,), dtype='uint8').realize(),
-                      Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
-    inputs = img_inputs + big_img_inputs
-    Device.default.synchronize()
-
-    st = time.perf_counter()
-    _ = update_img_jit(*inputs)
-    mt = time.perf_counter()
-    Device.default.synchronize()
-    et = time.perf_counter()
-    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
-
-  pkl_path = warp_pkl_path(cam_w, cam_h)
-  with open(pkl_path, "wb") as f:
-    pickle.dump(update_img_jit, f)
-  print(f"  Saved to {pkl_path}")
-
-  jit = pickle.load(open(pkl_path, "rb"))
-  jit(*inputs)
-
-
-def compile_dm_warp(cam_w, cam_h):
-  dm_w, dm_h = DM_INPUT_SIZE
-  _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
-
-  print(f"Compiling DM warp for {cam_w}x{cam_h}...")
-
-  warp_dm = make_warp_dm(cam_w, cam_h, dm_w, dm_h)
-  warp_dm_jit = TinyJit(warp_dm, prune=True)
-
-  new_frame_np = np.random.randint(0, 256, yuv_size, dtype=np.uint8)
-  for i in range(10):
-    inputs = [Tensor.from_blob(new_frame_np.ctypes.data, (yuv_size,), dtype='uint8').realize(),
-              Tensor(Tensor.randn(3, 3).mul(8).realize().numpy(), device='NPY')]
-    Device.default.synchronize()
-    st = time.perf_counter()
-    warp_dm_jit(*inputs)
-    mt = time.perf_counter()
-    Device.default.synchronize()
-    et = time.perf_counter()
-    print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
-
-  pkl_path = dm_warp_pkl_path(cam_w, cam_h)
-  with open(pkl_path, "wb") as f:
-    pickle.dump(warp_dm_jit, f)
-  print(f"  Saved to {pkl_path}")
-
-
-def run_and_save_pickle():
-  for cam_w, cam_h in CAMERA_CONFIGS:
-    compile_modeld_warp(cam_w, cam_h)
-    compile_dm_warp(cam_w, cam_h)
-
-
-if __name__ == "__main__":
-  run_and_save_pickle()

selfdrive/modeld/dmonitoringmodeld.py

@@ -1,12 +1,8 @@
 #!/usr/bin/env python3
 import os
-from openpilot.selfdrive.modeld.tinygrad_helpers import MODELS_DIR, set_tinygrad_backend_from_compiled_flags
+from openpilot.selfdrive.modeld.helpers import MODELS_DIR, CompileConfig, set_tinygrad_backend_from_compiled_flags
 set_tinygrad_backend_from_compiled_flags()
 
-# FIXME-SP: remove once we bump tg
-from openpilot.system.hardware import TICI
-os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
-
 from tinygrad.tensor import Tensor
 import time
 import pickle
@@ -32,7 +28,7 @@ class ModelState:
   inputs: dict[str, np.ndarray]
   output: np.ndarray
 
-  def __init__(self):
+  def __init__(self, cam_w: int, cam_h: int):
     with open(METADATA_PATH, 'rb') as f:
       model_metadata = pickle.load(f)
       self.input_shapes = model_metadata['input_shapes']
@@ -44,22 +40,18 @@ class ModelState:
 
     self.warp_inputs_np = {'transform': np.zeros((3,3), dtype=np.float32)}
     self.warp_inputs = {k: Tensor(v, device='NPY') for k,v in self.warp_inputs_np.items()}
-    self.frame_buf_params = None
+    self.frame_buf_params = get_nv12_info(cam_w, cam_h)
     self.tensor_inputs = {k: Tensor(v, device='NPY').realize() for k,v in self.numpy_inputs.items()}
     self._blob_cache : dict[int, Tensor] = {}
-    self.image_warp = None
     self.model_run = pickle.loads(read_file_chunked(str(MODEL_PKL_PATH)))
+    with open(CompileConfig(cam_w, cam_h, prefix='dm_', prepare_only=True).pkl_path, "rb") as f:
+      self.image_warp = pickle.load(f)
 
   def run(self, buf: VisionBuf, calib: np.ndarray, transform: np.ndarray) -> tuple[np.ndarray, float]:
     self.numpy_inputs['calib'][0,:] = calib
 
     t1 = time.perf_counter()
 
-    if self.image_warp is None:
-      self.frame_buf_params = get_nv12_info(buf.width, buf.height)
-      warp_path = MODELS_DIR / f'dm_warp_{buf.width}x{buf.height}_tinygrad.pkl'
-      with open(warp_path, "rb") as f:
-        self.image_warp = pickle.load(f)
     ptr = buf.data.ctypes.data
     # There is a ringbuffer of imgs, just cache tensors pointing to all of them
     if ptr not in self._blob_cache:
@@ -113,9 +105,6 @@ def get_driverstate_packet(model_output, frame_id: int, location_ts: int, exec_t
 def main():
   config_realtime_process(7, 5)
 
-  model = ModelState()
-  cloudlog.warning("models loaded, dmonitoringmodeld starting")
-
   cloudlog.warning("connecting to driver stream")
   vipc_client = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_DRIVER, True)
   while not vipc_client.connect(False):
@@ -123,6 +112,9 @@ def main():
   assert vipc_client.is_connected()
   cloudlog.warning(f"connected with buffer size: {vipc_client.buffer_len}")
 
+  model = ModelState(vipc_client.width, vipc_client.height)
+  cloudlog.warning("models loaded, dmonitoringmodeld starting")
+
   sm = SubMaster(["liveCalibration"])
   pm = PubMaster(["driverStateV2"])
 

selfdrive/modeld/get_model_metadata.py

@@ -7,6 +7,10 @@ from typing import Any
 
 from tinygrad.nn.onnx import OnnxPBParser
 
+def metadata_path_for(onnx_path) -> pathlib.Path:
+  p = pathlib.Path(onnx_path)
+  return p.parent / (p.stem + '_metadata.pkl')
+
 
 class MetadataOnnxPBParser(OnnxPBParser):
   def _parse_ModelProto(self) -> dict:
@@ -48,7 +52,7 @@ if __name__ == "__main__":
     'output_shapes': dict(get_name_and_shape(x) for x in model["graph"]["output"]),
   }
 
-  metadata_path = model_path.parent / (model_path.stem + '_metadata.pkl')
+  metadata_path = metadata_path_for(model_path)
   with open(metadata_path, 'wb') as f:
     pickle.dump(metadata, f)
 

selfdrive/modeld/helpers.py

@@ -0,0 +1,31 @@
+import json
+import os
+from dataclasses import dataclass
+from pathlib import Path
+
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+
+MODELS_DIR = Path(__file__).resolve().parent / 'models'
+COMPILED_FLAGS_PATH = MODELS_DIR / 'tg_compiled_flags.json'
+
+
+def set_tinygrad_backend_from_compiled_flags() -> None:
+  if os.path.isfile(COMPILED_FLAGS_PATH):
+    with open(COMPILED_FLAGS_PATH) as f:
+      os.environ['DEV'] = str(json.load(f)['DEV'])
+
+
+@dataclass
+class CompileConfig:
+  cam_w: int
+  cam_h: int
+  prepare_only: bool
+  prefix: str
+
+  @property
+  def pkl_path(self):
+    return str(MODELS_DIR / f'{self.prefix}{"warp_" if self.prepare_only else ""}{self.cam_w}x{self.cam_h}_tinygrad.pkl')
+
+  @property
+  def nv12(self):
+    return (self.cam_w, self.cam_h, *get_nv12_info(self.cam_w, self.cam_h))

selfdrive/modeld/modeld.py

@@ -1,12 +1,8 @@
 #!/usr/bin/env python3
 import os
-from openpilot.selfdrive.modeld.tinygrad_helpers import MODELS_DIR, set_tinygrad_backend_from_compiled_flags
+from openpilot.selfdrive.modeld.helpers import MODELS_DIR, CompileConfig, set_tinygrad_backend_from_compiled_flags
 set_tinygrad_backend_from_compiled_flags()
 
-# FIXME-SP: remove once we bump tg
-from openpilot.system.hardware import TICI
-os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
-
 USBGPU = "USBGPU" in os.environ
 if USBGPU:
   os.environ['DEV'] = 'AMD'
@@ -30,6 +26,7 @@ from openpilot.common.transformations.model import get_warp_matrix
 from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value, get_curvature_from_plan
 from openpilot.selfdrive.modeld.parse_model_outputs import Parser
+from openpilot.selfdrive.modeld.compile_modeld import make_input_queues
 from openpilot.selfdrive.modeld.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState
 from openpilot.common.file_chunker import read_file_chunked
 from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
@@ -41,17 +38,13 @@ from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
 PROCESS_NAME = "selfdrive.modeld.modeld"
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
 
-VISION_PKL_PATH = MODELS_DIR / 'driving_vision_tinygrad.pkl'
 VISION_METADATA_PATH = MODELS_DIR / 'driving_vision_metadata.pkl'
-POLICY_PKL_PATH = MODELS_DIR / 'driving_policy_tinygrad.pkl'
 POLICY_METADATA_PATH = MODELS_DIR / 'driving_policy_metadata.pkl'
 
 LAT_SMOOTH_SECONDS = 0.0
 LONG_SMOOTH_SECONDS = 0.3
 MIN_LAT_CONTROL_SPEED = 0.3
 
-IMG_QUEUE_SHAPE = (6*(ModelConstants.MODEL_RUN_FREQ//ModelConstants.MODEL_CONTEXT_FREQ + 1), 128, 256)
-assert IMG_QUEUE_SHAPE[0] == 30
 
 
 def get_action_from_model(model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
@@ -86,108 +79,39 @@ class FrameMeta:
     if vipc is not None:
       self.frame_id, self.timestamp_sof, self.timestamp_eof = vipc.frame_id, vipc.timestamp_sof, vipc.timestamp_eof
 
-class InputQueues:
-  def __init__ (self, model_fps, env_fps, n_frames_input):
-    assert env_fps % model_fps == 0
-    assert env_fps >= model_fps
-    self.model_fps = model_fps
-    self.env_fps = env_fps
-    self.n_frames_input = n_frames_input
-
-    self.dtypes = {}
-    self.shapes = {}
-    self.q = {}
-
-  def update_dtypes_and_shapes(self, input_dtypes, input_shapes) -> None:
-    self.dtypes.update(input_dtypes)
-    if self.env_fps == self.model_fps:
-      self.shapes.update(input_shapes)
-    else:
-      for k in input_shapes:
-        shape = list(input_shapes[k])
-        if 'img' in k:
-          n_channels = shape[1] // self.n_frames_input
-          shape[1] = (self.env_fps // self.model_fps + (self.n_frames_input - 1)) * n_channels
-        else:
-          shape[1] = (self.env_fps // self.model_fps) * shape[1]
-        self.shapes[k] = tuple(shape)
-
-  def reset(self) -> None:
-    self.q = {k: np.zeros(self.shapes[k], dtype=self.dtypes[k]) for k in self.dtypes.keys()}
-
-  def enqueue(self, inputs:dict[str, np.ndarray]) -> None:
-    for k in inputs.keys():
-      if inputs[k].dtype != self.dtypes[k]:
-        raise ValueError(f'supplied input <{k}({inputs[k].dtype})> has wrong dtype, expected {self.dtypes[k]}')
-      input_shape = list(self.shapes[k])
-      input_shape[1] = -1
-      single_input = inputs[k].reshape(tuple(input_shape))
-      sz = single_input.shape[1]
-      self.q[k][:,:-sz] = self.q[k][:,sz:]
-      self.q[k][:,-sz:] = single_input
-
-  def get(self, *names) -> dict[str, np.ndarray]:
-    if self.env_fps == self.model_fps:
-      return {k: self.q[k] for k in names}
-    else:
-      out = {}
-      for k in names:
-        shape = self.shapes[k]
-        if 'img' in k:
-          n_channels = shape[1] // (self.env_fps // self.model_fps + (self.n_frames_input - 1))
-          out[k] = np.concatenate([self.q[k][:, s:s+n_channels] for s in np.linspace(0, shape[1] - n_channels, self.n_frames_input, dtype=int)], axis=1)
-        elif 'pulse' in k:
-          # any pulse within interval counts
-          out[k] = self.q[k].reshape((shape[0], shape[1] * self.model_fps // self.env_fps, self.env_fps // self.model_fps, -1)).max(axis=2)
-        else:
-          idxs = np.arange(-1, -shape[1], -self.env_fps // self.model_fps)[::-1]
-          out[k] = self.q[k][:, idxs]
-      return out
 
 class ModelState(ModelStateBase):
-  inputs: dict[str, np.ndarray]
-  output: np.ndarray
   prev_desire: np.ndarray  # for tracking the rising edge of the pulse
 
-  def __init__(self):
+  def __init__(self, cam_w: int, cam_h: int):
     ModelStateBase.__init__(self)
     self.LAT_SMOOTH_SECONDS = LAT_SMOOTH_SECONDS
+
     with open(VISION_METADATA_PATH, 'rb') as f:
       vision_metadata = pickle.load(f)
       self.vision_input_shapes =  vision_metadata['input_shapes']
       self.vision_input_names = list(self.vision_input_shapes.keys())
       self.vision_output_slices = vision_metadata['output_slices']
-      vision_output_size = vision_metadata['output_shapes']['outputs'][1]
 
     with open(POLICY_METADATA_PATH, 'rb') as f:
       policy_metadata = pickle.load(f)
       self.policy_input_shapes =  policy_metadata['input_shapes']
       self.policy_output_slices = policy_metadata['output_slices']
-      policy_output_size = policy_metadata['output_shapes']['outputs'][1]
 
     self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
 
-    # policy inputs
+    self.frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
-    self.numpy_inputs = {k: np.zeros(self.policy_input_shapes[k], dtype=np.float32) for k in self.policy_input_shapes}
+    self.input_queues, self.npy = make_input_queues(self.vision_input_shapes, self.policy_input_shapes, self.frame_skip)
-    self.full_input_queues = InputQueues(ModelConstants.MODEL_CONTEXT_FREQ, ModelConstants.MODEL_RUN_FREQ, ModelConstants.N_FRAMES)
-    for k in ['desire_pulse', 'features_buffer']:
-      self.full_input_queues.update_dtypes_and_shapes({k: self.numpy_inputs[k].dtype}, {k: self.numpy_inputs[k].shape})
-    self.full_input_queues.reset()
-
-    self.img_queues = {'img': Tensor.zeros(IMG_QUEUE_SHAPE, dtype='uint8').contiguous().realize(),
-                       'big_img': Tensor.zeros(IMG_QUEUE_SHAPE, dtype='uint8').contiguous().realize()}
     self.full_frames : dict[str, Tensor] = {}
     self._blob_cache : dict[int, Tensor] = {}
-    self.transforms_np = {k: np.zeros((3,3), dtype=np.float32) for k in self.img_queues}
-    self.transforms = {k: Tensor(v, device='NPY').realize() for k, v in self.transforms_np.items()}
-    self.vision_output = np.zeros(vision_output_size, dtype=np.float32)
-    self.policy_inputs = {k: Tensor(v, device='NPY').realize() for k,v in self.numpy_inputs.items()}
-    self.policy_output = np.zeros(policy_output_size, dtype=np.float32)
     self.parser = Parser()
-    self.frame_buf_params : dict[str, tuple[int, int, int, int]] = {}
+    self.frame_buf_params = {k: get_nv12_info(cam_w, cam_h) for k in ('img', 'big_img')}
-    self.update_imgs = None
+    self.run_policy = pickle.loads(read_file_chunked(CompileConfig(cam_w, cam_h, prefix='driving_', prepare_only=False).pkl_path))
-    self.vision_run = pickle.loads(read_file_chunked(str(VISION_PKL_PATH)))
+    self.warp_enqueue = pickle.loads(read_file_chunked(CompileConfig(cam_w, cam_h, prefix='driving_', prepare_only=True).pkl_path))
-    self.policy_run = pickle.loads(read_file_chunked(str(POLICY_PKL_PATH)))
+    self.warp_enqueue(
+      **self.input_queues,
+      frame=Tensor.zeros(self.frame_buf_params['img'][3], dtype='uint8').contiguous().realize(),
+      big_frame=Tensor.zeros(self.frame_buf_params['big_img'][3], dtype='uint8').contiguous().realize())
 
   def slice_outputs(self, model_outputs: np.ndarray, output_slices: dict[str, slice]) -> dict[str, np.ndarray]:
     parsed_model_outputs = {k: model_outputs[np.newaxis, v] for k,v in output_slices.items()}
@@ -195,18 +119,6 @@ class ModelState(ModelStateBase):
 
   def run(self, bufs: dict[str, VisionBuf], transforms: dict[str, np.ndarray],
                 inputs: dict[str, np.ndarray], prepare_only: bool) -> dict[str, np.ndarray] | None:
-    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
-    inputs['desire_pulse'][0] = 0
-    new_desire = np.where(inputs['desire_pulse'] - self.prev_desire > .99, inputs['desire_pulse'], 0)
-    self.prev_desire[:] = inputs['desire_pulse']
-    if self.update_imgs is None:
-      for key in bufs.keys():
-        w, h = bufs[key].width, bufs[key].height
-        self.frame_buf_params[key] = get_nv12_info(w, h)
-      warp_path = MODELS_DIR / f'warp_{w}x{h}_tinygrad.pkl'
-      with open(warp_path, "rb") as f:
-        self.update_imgs = pickle.load(f)
-
     for key in bufs.keys():
       ptr = bufs[key].data.ctypes.data
       yuv_size = self.frame_buf_params[key][3]
@@ -215,30 +127,31 @@ class ModelState(ModelStateBase):
       if cache_key not in self._blob_cache:
         self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8')
       self.full_frames[key] = self._blob_cache[cache_key]
-    for key in bufs.keys():
-      self.transforms_np[key][:,:] = transforms[key][:,:]
 
-    out = self.update_imgs(self.img_queues['img'], self.full_frames['img'], self.transforms['img'],
+    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
-                           self.img_queues['big_img'], self.full_frames['big_img'], self.transforms['big_img'])
+    inputs['desire_pulse'][0] = 0
-    vision_inputs = {'img': out[0], 'big_img': out[1]}
+    self.npy['desire'][:] = np.where(inputs['desire_pulse'] - self.prev_desire > .99, inputs['desire_pulse'], 0)
+    self.prev_desire[:] = inputs['desire_pulse']
+    self.npy['traffic_convention'][:] = inputs['traffic_convention']
+    self.npy['tfm'][:,:] = transforms['img'][:,:]
+    self.npy['big_tfm'][:,:] = transforms['big_img'][:,:]
 
     if prepare_only:
+      self.warp_enqueue(**self.input_queues, frame=self.full_frames['img'], big_frame=self.full_frames['big_img'])
       return None
 
-    self.vision_output = self.vision_run(**vision_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    vision_output, policy_output = self.run_policy(
-    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(self.vision_output, self.vision_output_slices))
+      **self.input_queues, frame=self.full_frames['img'], big_frame=self.full_frames['big_img']
+    )
 
-    self.full_input_queues.enqueue({'features_buffer': vision_outputs_dict['hidden_state'], 'desire_pulse': new_desire})
+    vision_output = vision_output.numpy().flatten()
-    for k in ['desire_pulse', 'features_buffer']:
+    policy_output = policy_output.numpy().flatten()
-      self.numpy_inputs[k][:] = self.full_input_queues.get(k)[k]
+    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(vision_output, self.vision_output_slices))
-    self.numpy_inputs['traffic_convention'][:] = inputs['traffic_convention']
+    policy_outputs_dict = self.parser.parse_policy_outputs(self.slice_outputs(policy_output, self.policy_output_slices))
-
-    self.policy_output = self.policy_run(**self.policy_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
-    policy_outputs_dict = self.parser.parse_policy_outputs(self.slice_outputs(self.policy_output, self.policy_output_slices))
     combined_outputs_dict = {**vision_outputs_dict, **policy_outputs_dict}
-    if SEND_RAW_PRED:
-      combined_outputs_dict['raw_pred'] = np.concatenate([self.vision_output.copy(), self.policy_output.copy()])
 
+    if SEND_RAW_PRED:
+      combined_outputs_dict['raw_pred'] = np.concatenate([vision_output.copy(), policy_output.copy()])
     return combined_outputs_dict
 
 
@@ -250,11 +163,6 @@ def main(demo=False):
     # also need to move the aux USB interrupts for good timings
     config_realtime_process(7, 54)
 
-  st = time.monotonic()
-  cloudlog.warning("loading model")
-  model = ModelState()
-  cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
-
   # visionipc clients
   while True:
     available_streams = VisionIpcClient.available_streams("camerad", block=False)
@@ -278,6 +186,11 @@ def main(demo=False):
   if use_extra_client:
     cloudlog.warning(f"connected extra cam with buffer size: {vipc_client_extra.buffer_len} ({vipc_client_extra.width} x {vipc_client_extra.height})")
 
+  st = time.monotonic()
+  cloudlog.warning("loading model")
+  model = ModelState(vipc_client_main.width, vipc_client_main.height)
+  cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
+
   # messaging
   pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"])
   sm = SubMaster(["deviceState", "carState", "roadCameraState", "liveCalibration", "driverMonitoringState", "carControl", "liveDelay"])

selfdrive/modeld/tinygrad_helpers.py

@@ -1,12 +0,0 @@
-import json
-import os
-from pathlib import Path
-
-MODELS_DIR = Path(__file__).parent / 'models'
-COMPILED_FLAGS_PATH = MODELS_DIR / 'tg_compiled_flags.json'
-
-
-def set_tinygrad_backend_from_compiled_flags() -> None:
-  if os.path.isfile(COMPILED_FLAGS_PATH):
-    with open(COMPILED_FLAGS_PATH) as f:
-      os.environ['DEV'] = str(json.load(f)['DEV'])

sunnypilot/modeld_v2/SConscript

@@ -1,5 +1,6 @@
 import os
 import glob
+from tinygrad import Device
 
 Import('env', 'arch')
 lenv = env.Clone()
@@ -21,10 +22,19 @@ if PC:
   if outputs:
     lenv.Command(outputs, inputs, cmd)
 
-tg_flags = {
+available = set(Device.get_available_devices())
-  'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0',
+if 'CUDA' in available:
-  'Darwin': f'DEV=CPU THREADS=0 HOME={os.path.expanduser("~")}',
+  tg_backend = 'CUDA'
-}.get(arch, 'DEV=CPU CPU_LLVM=1 THREADS=0')
+  tg_flags = f'DEV={tg_backend}'
+elif 'QCOM' in available:
+  tg_backend = 'QCOM'
+  tg_flags = f'DEV={tg_backend} FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0'
+else:
+  tg_backend = 'CPU' if arch == 'Darwin' else 'CPU:LLVM'
+  # THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
+  tg_flags = f'DEV={tg_backend} THREADS=0'
+
+mac_brew_string = f'HOME={os.path.expanduser("~")}' if arch == 'Darwin' else ''
 
 image_flag = {
   'larch64': 'IMAGE=2',
@@ -38,7 +48,7 @@ def tg_compile(flags, model_name):
   return lenv.Command(
     out,
     [fn + ".onnx"] + tinygrad_files,
-    f'{pythonpath_string} {flags} {image_flag} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {out}'
+    f'{pythonpath_string} {tg_flags} {mac_brew_string} {image_flag} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {out}'
   )
 
 # Compile models
@@ -46,9 +56,9 @@ for model_name in ['supercombo', 'driving_vision', 'driving_off_policy', 'drivin
   if File(f"models/{model_name}.onnx").exists():
     tg_compile(tg_flags, model_name)
 
-script_files = [File("warp.py"), File(Dir("#selfdrive/modeld").File("compile_warp.py").abspath)]
+script_files = [File("warp.py")]
 pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + ':' + env.Dir("#").abspath + '"'
-compile_warp_cmd = f'{pythonpath_string} {tg_flags} python3 -m sunnypilot.modeld_v2.warp'
+compile_warp_cmd = f'{pythonpath_string} {tg_flags} {mac_brew_string} {image_flag} python3 -m sunnypilot.modeld_v2.warp'
 
 from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
 warp_targets = []

sunnypilot/modeld_v2/modeld.py

@@ -129,8 +129,7 @@ class ModelState(ModelStateBase):
         self.numpy_inputs[key][:] = inputs[key]
 
     imgs_tensors = self.warp.process(bufs, transforms)
-    for name, tensor in imgs_tensors.items():
+    self.model_runner.update_vision_inputs(imgs_tensors)
-      self.model_runner.inputs[name] = tensor
     self.model_runner.prepare_inputs(self.numpy_inputs)
 
     if prepare_only:

sunnypilot/modeld_v2/tests/test_warp.py

@@ -2,8 +2,11 @@ import os
 os.environ['DEV'] = 'CPU'
 import pytest
 import numpy as np
-from openpilot.selfdrive.modeld.compile_warp import get_nv12_info, CAMERA_CONFIGS
+from openpilot.sunnypilot.modeld_v2.warp import CAMERA_CONFIGS
-from openpilot.sunnypilot.modeld_v2.warp import Warp, MODEL_W, MODEL_H
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.sunnypilot.modeld_v2.warp import Warp
+from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE
+MODEL_W, MODEL_H = MEDMODEL_INPUT_SIZE
 
 VISION_NAME_PAIRS = [  # needed to account for supercombos input_imgs
   ('img', 'big_img'),

sunnypilot/modeld_v2/warp.py

@@ -6,29 +6,128 @@ from tinygrad.tensor import Tensor
 from tinygrad.engine.jit import TinyJit
 from tinygrad.device import Device
 
+from typing import NamedTuple
+# https://github.com/tinygrad/tinygrad/issues/15682
+from tinygrad.uop.ops import UOp, Ops
+_orig = UOp.__reduce__
+UOp.__reduce__ = lambda self: (UOp.unique, ()) if self.op is Ops.UNIQUE else _orig(self)
+
+from tinygrad.helpers import Context
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-from openpilot.selfdrive.modeld.compile_warp import (
+from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
-  CAMERA_CONFIGS, MEDMODEL_INPUT_SIZE, make_frame_prepare, make_update_both_imgs,
+
-  warp_pkl_path,
+class NV12Frame(NamedTuple):
-)
+  cam_w: int
+  cam_h: int
+  stride: int
+  y_height: int
+  uv_height: int
+  size: int
+
+UV_SCALE_MATRIX = np.array([[0.5, 0, 0], [0, 0.5, 0], [0, 0, 1]], dtype=np.float32)
+UV_SCALE_MATRIX_INV = np.linalg.inv(UV_SCALE_MATRIX)
+
+CAMERA_CONFIGS = [
+  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
+  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
+]
+from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE
 
 MODELS_DIR = Path(__file__).parent / 'models'
-MODEL_W, MODEL_H = MEDMODEL_INPUT_SIZE
+
 UPSTREAM_BUFFER_LENGTH = 5
 
+def warp_pkl_path(cam_w, cam_h):
+  return MODELS_DIR / f'warp_{cam_w}x{cam_h}_tinygrad.pkl'
+
+def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad):
+  w_dst, h_dst = dst_shape
+  h_src, w_src = src_shape
+
+  x = Tensor.arange(w_dst).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
+  y = Tensor.arange(h_dst).reshape(h_dst, 1).expand(h_dst, w_dst).reshape(-1)
+
+  # inline 3x3 matmul as elementwise to avoid reduce op (enables fusion with gather)
+  src_x = M_inv[0, 0] * x + M_inv[0, 1] * y + M_inv[0, 2]
+  src_y = M_inv[1, 0] * x + M_inv[1, 1] * y + M_inv[1, 2]
+  src_w = M_inv[2, 0] * x + M_inv[2, 1] * y + M_inv[2, 2]
+
+  src_x = src_x / src_w
+  src_y = src_y / src_w
+
+  x_nn_clipped = Tensor.round(src_x).clip(0, w_src - 1).cast('int')
+  y_nn_clipped = Tensor.round(src_y).clip(0, h_src - 1).cast('int')
+  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
+
+  return src_flat[idx]
+
+def frames_to_tensor(frames, model_w, model_h):
+  H = (frames.shape[0] * 2) // 3
+  W = frames.shape[1]
+  in_img1 = Tensor.cat(frames[0:H:2, 0::2],
+                       frames[1:H:2, 0::2],
+                       frames[0:H:2, 1::2],
+                       frames[1:H:2, 1::2],
+                       frames[H:H+H//4].reshape((H//2, W//2)),
+                       frames[H+H//4:H+H//2].reshape((H//2, W//2)), dim=0).reshape((6, H//2, W//2))
+  return in_img1
+
+def make_frame_prepare(cam_w, cam_h, model_w, model_h):
+  stride, y_height, uv_height, _ = get_nv12_info(cam_w, cam_h)
+  uv_offset = stride * y_height
+  stride_pad = stride - cam_w
+
+  def frame_prepare_tinygrad(input_frame, M_inv):
+    # UV_SCALE @ M_inv @ UV_SCALE_INV simplifies to elementwise scaling
+    M_inv_uv = M_inv * Tensor([[1.0, 1.0, 0.5], [1.0, 1.0, 0.5], [2.0, 2.0, 1.0]])
+    # deinterleave NV12 UV plane (UVUV... -> separate U, V)
+    uv = input_frame[uv_offset:uv_offset + uv_height * stride].reshape(uv_height, stride)
+    with Context(SPLIT_REDUCEOP=0):
+      y = warp_perspective_tinygrad(input_frame[:cam_h*stride],
+                                    M_inv, (model_w, model_h),
+                                    (cam_h, cam_w), stride_pad).realize()
+      u = warp_perspective_tinygrad(uv[:cam_h//2, :cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+      v = warp_perspective_tinygrad(uv[:cam_h//2, 1:cam_w:2].flatten(),
+                                    M_inv_uv, (model_w//2, model_h//2),
+                                    (cam_h//2, cam_w//2), 0).realize()
+    yuv = y.cat(u).cat(v).reshape((model_h * 3 // 2, model_w))
+    tensor = frames_to_tensor(yuv, model_w, model_h)
+    return tensor
+  return frame_prepare_tinygrad
+
+def make_update_img_input(frame_prepare, model_w, model_h):
+  def update_img_input_tinygrad(tensor, frame, M_inv):
+    M_inv = M_inv.to(Device.DEFAULT)
+    new_img = frame_prepare(frame, M_inv)
+    tensor.assign(tensor[6:].cat(new_img, dim=0).contiguous())
+    return tensor, Tensor.cat(tensor[:6], tensor[-6:], dim=0).contiguous().reshape(1, 12, model_h//2, model_w//2)
+  return update_img_input_tinygrad
+
+def make_update_both_imgs(frame_prepare, model_w, model_h):
+  update_img = make_update_img_input(frame_prepare, model_w, model_h)
+
+  def update_both_imgs_tinygrad(calib_img_buffer, new_img, M_inv,
+                                calib_big_img_buffer, new_big_img, M_inv_big):
+    r1, r2 = update_img(calib_img_buffer, new_img, M_inv)
+    w1, w2 = update_img(calib_big_img_buffer, new_big_img, M_inv_big)
+    return r1, r2, w1, w2
+  return update_both_imgs_tinygrad
+
 
 def v2_warp_pkl_path(cam_w, cam_h, buffer_length):
   return MODELS_DIR / f'warp_{cam_w}x{cam_h}_b{buffer_length}_tinygrad.pkl'
 
 
-def compile_v2_warp(cam_w, cam_h, buffer_length):
+def compile_v2_warp(cam_w, cam_h, buffer_length, model_w=MEDMODEL_INPUT_SIZE[0], model_h=MEDMODEL_INPUT_SIZE[1], pkl_path=None):
   _, _, _, yuv_size = get_nv12_info(cam_w, cam_h)
-  img_buffer_shape = (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+  img_buffer_shape = (buffer_length * 6, model_h // 2, model_w // 2)
 
   print(f"Compiling v2 warp for {cam_w}x{cam_h} buffer_length={buffer_length}...")
 
-  frame_prepare = make_frame_prepare(cam_w, cam_h, MODEL_W, MODEL_H)
+  frame_prepare = make_frame_prepare(cam_w, cam_h, model_w, model_h)
-  update_both_imgs = make_update_both_imgs(frame_prepare, MODEL_W, MODEL_H)
+  update_both_imgs = make_update_both_imgs(frame_prepare, model_w, model_h)
   update_img_jit = TinyJit(update_both_imgs, prune=True)
 
   full_buffer = Tensor.zeros(img_buffer_shape, dtype='uint8').contiguous().realize()
@@ -46,25 +145,25 @@ def compile_v2_warp(cam_w, cam_h, buffer_length):
     Device.default.synchronize()
 
     st = time.perf_counter()
-    _ = update_img_jit(*inputs)
+    update_img_jit(*inputs)
     mt = time.perf_counter()
     Device.default.synchronize()
     et = time.perf_counter()
     print(f"  [{i+1}/10] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
 
-  pkl_path = v2_warp_pkl_path(cam_w, cam_h, buffer_length)
+  if pkl_path is None:
+    pkl_path = v2_warp_pkl_path(cam_w, cam_h, buffer_length)
   with open(pkl_path, "wb") as f:
     pickle.dump(update_img_jit, f)
   print(f"  Saved to {pkl_path}")
 
-  jit = pickle.load(open(pkl_path, "rb"))
-  jit(*inputs)
-
 
 class Warp:
-  def __init__(self, buffer_length=2):
+  def __init__(self, buffer_length=2, model_w=MEDMODEL_INPUT_SIZE[0], model_h=MEDMODEL_INPUT_SIZE[1]):
     self.buffer_length = buffer_length
-    self.img_buffer_shape = (buffer_length * 6, MODEL_H // 2, MODEL_W // 2)
+    self.model_w = model_w
+    self.model_h = model_h
+    self.img_buffer_shape = (buffer_length * 6, model_h // 2, model_w // 2)
 
     self.jit_cache = {}
     self.full_buffers = {k: Tensor.zeros(self.img_buffer_shape, dtype='uint8').contiguous().realize() for k in ['img', 'big_img']}
@@ -92,8 +191,8 @@ class Warp:
           with open(upstream_pkl, 'rb') as f:
             self.jit_cache[key] = pickle.load(f)
       if key not in self.jit_cache:
-        frame_prepare = make_frame_prepare(cam_w, cam_h, MODEL_W, MODEL_H)
+        frame_prepare = make_frame_prepare(cam_w, cam_h, self.model_w, self.model_h)
-        update_both_imgs = make_update_both_imgs(frame_prepare, MODEL_W, MODEL_H)
+        update_both_imgs = make_update_both_imgs(frame_prepare, self.model_w, self.model_h)
         self.jit_cache[key] = TinyJit(update_both_imgs, prune=True)
 
     if key not in self._nv12_cache:
@@ -107,7 +206,7 @@ class Warp:
     if wide_ptr not in self._blob_cache:
       self._blob_cache[wide_ptr] = Tensor.from_blob(wide_ptr, (yuv_size,), dtype='uint8')
     road_blob = self._blob_cache[road_ptr]
-    wide_blob = self._blob_cache[wide_ptr] if wide_ptr != road_ptr else Tensor.from_blob(wide_ptr, (yuv_size,), dtype='uint8')
+    wide_blob = self._blob_cache[wide_ptr]
     np.copyto(self.transforms_np['img'], transforms[road].reshape(3, 3))
     np.copyto(self.transforms_np['big_img'], transforms[wide].reshape(3, 3))
 
@@ -116,13 +215,11 @@ class Warp:
       self.full_buffers['img'], road_blob, self.transforms['img'],
       self.full_buffers['big_img'], wide_blob, self.transforms['big_img'],
     )
-    out_road = res[0].realize()
+    return {road: res[1].realize(), wide: res[3].realize()}
-    out_wide = res[1].realize()
-
-    return {road: out_road, wide: out_wide}
 
 
 if __name__ == "__main__":
   for cam_w, cam_h in CAMERA_CONFIGS:
+    compile_v2_warp(cam_w, cam_h, 5, pkl_path=warp_pkl_path(cam_w, cam_h))
     for bl in [2, 5]:
       compile_v2_warp(cam_w, cam_h, bl)

sunnypilot/models/fetcher.py

@@ -116,7 +116,7 @@ class ModelCache:
 
 class ModelFetcher:
   """Handles fetching and caching of model data from remote source"""
-  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v16.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v18.json"
 
   def __init__(self, params: Params):
     self.params = params

sunnypilot/models/runners/model_runner.py

@@ -132,6 +132,11 @@ class ModelRunner(ModularRunner):
       return list(self._model_data.input_shapes.keys())
     raise ValueError("Model data is not available. Ensure the model is loaded correctly.")
 
+  def update_vision_inputs(self, vision_inputs: dict) -> None:
+    """Updates the vision inputs in the runner."""
+    for name, tensor in vision_inputs.items():
+      self.inputs[name] = tensor
+
   @abstractmethod
   def prepare_inputs(self, numpy_inputs: NumpyDict) -> dict:
     """

sunnypilot/models/runners/tinygrad/tinygrad_runner.py

@@ -46,14 +46,13 @@ class TinygradRunner(ModelRunner, SupercomboTinygrad, PolicyTinygrad, VisionTiny
         assert "/dev/kgsl-3d0" not in str(e), "Model was built on C3 or C3X, but is being loaded on PC"
         raise
 
-    # Map input names to their required dtype and device from the loaded model
     self.input_to_dtype = {}
     self.input_to_device = {}
     for idx, name in enumerate(self.model_run.captured.expected_names):
       info = self.model_run.captured.expected_input_info[idx]
-      self.input_to_dtype[name] = info[2]  # dtype
+      self.input_to_dtype[name] = info[2]
-      self.input_to_device[name] = info[3]  # device
+      self.input_to_device[name] = info[3]
-    self._policy_cached = False
+      self.inputs[name] = Tensor.zeros(*self.input_shapes[name], dtype=info[2], device=info[3]).realize()
 
   @property
   def vision_input_names(self) -> list[str]:
@@ -62,22 +61,23 @@ class TinygradRunner(ModelRunner, SupercomboTinygrad, PolicyTinygrad, VisionTiny
 
 
   def prepare_policy_inputs(self, numpy_inputs: NumpyDict):
-    if not self._policy_cached:
+    for key, value in numpy_inputs.items():
-      for key, value in numpy_inputs.items():
+      if key in self.inputs:
-        self.inputs[key] = Tensor(value, device='NPY').realize()
+        self.inputs[key].assign(Tensor(value, device=self.inputs[key].device))
-      self._policy_cached = True
 
   def prepare_inputs(self, numpy_inputs: NumpyDict) -> dict:
     """Prepares all vision and policy inputs for the model."""
     self.prepare_policy_inputs(numpy_inputs)
-    for key in self.vision_input_names:
-      if key in self.inputs:
-        self.inputs[key] = self.inputs[key].cast(self.input_to_dtype[key])
     return self.inputs
 
+  def update_vision_inputs(self, vision_inputs: dict[str, Tensor]):
+    for name, tensor in vision_inputs.items():
+      if name in self.inputs:
+        self.inputs[name].assign(tensor)
+
   def _run_model(self) -> NumpyDict:
     """Runs the Tinygrad model inference and parses the outputs."""
-    outputs = self.model_run(**self.inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    outputs = self.model_run(**self.inputs).numpy().flatten()
     return self._parse_outputs(outputs)
 
   def _parse_outputs(self, model_outputs: np.ndarray) -> NumpyDict: