Use a precompiled eGPU driving model (#38930)

* Ship precompiled eGPU model and camera warps

Compile f78ed37d-afad-4dbc-8050-40ea885eedde/12864 through xx/ml_tools/openpilot_compile using the pinned tinygrad version.

* Precompile the existing master driving model

Use the unchanged master ONNX (SHA-256 6fee5937923c74848df4a63f6239eb6331c6274dd4bdb7a5d6ec0388a8b543d5) instead of updating the trained model.

* Compile camera warps on device

* Remove obsolete ONNX chunking and big model build check

* Chunk model artifacts only during release packaging

* Require model and camera warps for Chestnut readiness

* Recompile precompiled CPU helpers for the runtime host

* Ship the eGPU model with an ARM submission helper

* Exempt model pickles from the build product size limit
This commit is contained in:
Harald Schäfer
2026-09-16 08:13:45 -07:00
committed by GitHub
parent 81ae1a2e2d
commit 6080cc6168
17 changed files with 39 additions and 72 deletions
+1
View File
@@ -3,6 +3,7 @@
# to move existing files into LFS:
# git add --renormalize .
*.onnx filter=lfs diff=lfs merge=lfs -text
openpilot/selfdrive/modeld/models/big_driving_tinygrad.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
+1 -1
View File
@@ -18,7 +18,7 @@ concurrency:
env:
GIT_CONFIG_COUNT: 1
GIT_CONFIG_KEY_0: lfs.fetchexclude
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
jobs:
docs:
+1 -1
View File
@@ -7,7 +7,7 @@ on:
env:
GIT_CONFIG_COUNT: 1
GIT_CONFIG_KEY_0: lfs.fetchexclude
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
jobs:
build_masterci:
+1 -1
View File
@@ -11,7 +11,7 @@ env:
PYTHONPATH: ${{ github.workspace }}
GIT_CONFIG_COUNT: 1
GIT_CONFIG_KEY_0: lfs.fetchexclude
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
jobs:
package_updates:
+1 -1
View File
@@ -22,7 +22,7 @@ env:
PYTHONPATH: ${{ github.workspace }}
GIT_CONFIG_COUNT: 1
GIT_CONFIG_KEY_0: lfs.fetchexclude
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx
GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
jobs:
build_release:
+1
View File
@@ -50,6 +50,7 @@ st[0-9A-Za-z][0-9A-Za-z][0-9A-Za-z][0-9A-Za-z][0-9A-Za-z][0-9A-Za-z]
*.stats
*.pkl
*.pkl*
!openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
config.json
compile_commands.json
compare_runtime*.html
+2
View File
@@ -342,6 +342,8 @@ AddPostAction(BUILD_TARGETS or [Dir('.')], prune_cache_dir)
def check_build_product_size(target, source, env):
limit = 50 * 1024 * 1024 # GitHub max size
for t in target:
if str(t).endswith('.pkl'): # chunked during release packaging
continue
if hasattr(t, 'isfile') and t.isfile() and (size := os.path.getsize(t.abspath)) > limit:
raise SCons.Errors.UserError(f"{t} is {size / (1024 * 1024):.1f} MiB, exceeding the {limit / (1024 * 1024):.1f} MiB limit")
if not GetOption('extras'):
+3 -11
View File
@@ -32,14 +32,6 @@ def chunk_file(path, targets):
Path(manifest_path).write_text(str(len(chunk_paths)))
os.remove(path)
def get_existing_chunks(path):
if os.path.isfile(path):
return [path]
if os.path.isfile(manifest := get_manifest_path(path)):
num_chunks = int(Path(manifest).read_text().strip())
return _chunk_paths(path, num_chunks)
raise FileNotFoundError(path)
class ChunkStream(io.RawIOBase):
def __init__(self, paths):
self._paths = iter(paths)
@@ -67,11 +59,11 @@ class ChunkStream(io.RawIOBase):
def open_file_chunked(path):
manifest_path = get_manifest_path(path)
if os.path.isfile(manifest_path):
if os.path.isfile(path):
paths = [path]
elif os.path.isfile(manifest_path):
num_chunks = int(Path(manifest_path).read_text().strip())
paths = [get_chunk_name(path, i, num_chunks) for i in range(num_chunks)]
elif os.path.isfile(path):
paths = [path]
else:
raise FileNotFoundError(path)
return io.BufferedReader(ChunkStream(paths))
+13 -42
View File
@@ -1,18 +1,14 @@
import glob
import os
import shutil
import tempfile
import time
from SCons.Script import Action, Value
from openpilot.common.file_chunker import chunk_file, get_chunk_targets, get_existing_chunks, open_file_chunked
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.helpers import chestnut_present, modeld_pkl_path
from openpilot.selfdrive.modeld.helpers import chestnut_present
from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
Import('env', 'arch')
chunker_file = File("#openpilot/common/file_chunker.py")
lenv = env.Clone()
lenv.PrependENVPath('PYTHONPATH', Dir('#tinygrad_repo').abspath)
@@ -20,11 +16,6 @@ tinygrad_root = env.Dir("#").abspath
tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "/**", recursive=True, root_dir=tinygrad_root)
if 'pycache' not in x and os.path.isfile(os.path.join(tinygrad_root, x))]
def estimate_pickle_max_size(onnx_size):
# QCOM programs for models with spatial recurrent features can approach 2x
# the ONNX size. Overestimating only adds an empty trailing chunk.
return 2.0 * onnx_size + 10 * 1024 * 1024
camera_configs = [(c.width, c.height) for c in (_ar_ox_fisheye, _os_fisheye)]
if arch == 'comma_arm64':
@@ -47,7 +38,7 @@ compiler = Dir('#tinygrad_repo/examples/openpilot').abspath
# CPU 7 is isolated with isolcpus on AGNOS, so explicitly pin the compiler to it.
taskset = 'taskset -c 7 ' if arch == 'comma_arm64' else ''
def chestnut_action(command, pkl=None, chunks=()):
def chestnut_action(command):
def do_compile(target, source, env):
from openpilot.system.hardware.chestnut.flash import link_up
# chestnut can enumerate before its PCIe link is up due to varying 12V power behavior across cars
@@ -56,47 +47,27 @@ def chestnut_action(command, pkl=None, chunks=()):
break
time.sleep(1)
else:
print("Chestnut not ready, skipping big model build")
print("Chestnut not ready, skipping warp build")
return
if ret := env.Execute(command):
return ret
if chunks:
chunk_file(pkl, chunks)
return env.Execute(command)
return Action(do_compile, " [CHESTNUT] $TARGET")
def compile_model(onnx_path, pkl_path, flags, chestnut=False):
def compile_model(onnx_path, pkl_path):
onnx_path, target_pkl_path = File(onnx_path).abspath, File(pkl_path).abspath
onnx_deps = get_existing_chunks(onnx_path)
cmd = (f'{flags} {mac_brew_string} {taskset}python3 "{compiler}/compile_onnx.py" '
f'"{{onnx}}" "{target_pkl_path}" --device-input "*" --out-of-band --benchmark-runs 1')
def do_compile(target, source, env):
if os.path.isfile(onnx_path):
return env.Execute(cmd.format(onnx=onnx_path))
# TODO: Remove ONNX chunk reassembly once models are precompiled.
with tempfile.NamedTemporaryFile(dir=os.path.dirname(onnx_path), suffix='.onnx') as tmp, open_file_chunked(onnx_path) as src:
shutil.copyfileobj(src, tmp)
tmp.flush()
return env.Execute(cmd.format(onnx=tmp.name))
compile_action = Action(do_compile, " [ONNX] $TARGET")
onnx_sizes_sum = sum(os.path.getsize(f) for f in onnx_deps)
chunk_targets = get_chunk_targets(target_pkl_path, estimate_pickle_max_size(onnx_sizes_sum))
def do_chunk(target, source, env, pkl=target_pkl_path, chunks=chunk_targets):
chunk_file(pkl, chunks)
actions = chestnut_action(compile_action, target_pkl_path, chunk_targets) if chestnut else [compile_action, Action(do_chunk, " [CHUNK] $TARGET")]
node = lenv.Command(
chunk_targets,
tinygrad_files + onnx_deps + [Value(cmd), Value(chunk_targets), chunker_file],
actions,
cmd = (f'{tg_flags} {mac_brew_string} {taskset}python3 "{compiler}/compile_onnx.py" '
f'"{onnx_path}" "{target_pkl_path}" --device-input "*" --out-of-band --benchmark-runs 1')
lenv.Command(
target_pkl_path,
tinygrad_files + [onnx_path, Value(cmd)],
Action(cmd, " [ONNX] $TARGET"),
)
if chestnut:
lenv.SideEffect(chestnut_lock, node)
compile_model('models/dmonitoring_model.onnx', 'models/dmonitoring_model_tinygrad.pkl', tg_flags)
compile_model('models/dmonitoring_model.onnx', 'models/dmonitoring_model_tinygrad.pkl')
compile_model('models/driving_supercombo.onnx', 'models/driving_tinygrad.pkl')
model_w, model_h = MEDMODEL_INPUT_SIZE
for chestnut in [False, True] if CHESTNUT else [False]:
file_prefix, cmd_flags = ('big_', chestnut_tg_flags) if chestnut else ('', tg_flags)
compile_model(f'models/{file_prefix}driving_supercombo.onnx', modeld_pkl_path(chestnut), cmd_flags, chestnut)
for cam_w, cam_h in camera_configs:
warp_pkl_path = File(f"models/{file_prefix}driving_warp_{cam_w}x{cam_h}_tinygrad.pkl").abspath
stride, y_height, uv_height, _ = get_nv12_info(cam_w, cam_h)
+3 -1
View File
@@ -35,4 +35,6 @@ def chestnut_present() -> bool:
return False
def chestnut_compiled() -> bool:
return Path(get_manifest_path(modeld_pkl_path(chestnut=True))).is_file()
path = modeld_pkl_path(chestnut=True)
return (path.is_file() or Path(get_manifest_path(path)).is_file()) and all(
(MODELS_DIR / f'big_driving_warp_{size}_tinygrad.pkl').is_file() for size in ('1344x760', '1928x1208'))
@@ -1,3 +0,0 @@
version https://git-lfs.github.com/spec/v1
oid sha256:6fee5937923c74848df4a63f6239eb6331c6274dd4bdb7a5d6ec0388a8b543d5
size 766018462
@@ -0,0 +1,3 @@
version https://git-lfs.github.com/spec/v1
oid sha256:76cc0a9bc3af7a318889483dcbe126337f8d338f5abcbe664a8988c9b18b6639
size 776634338
-3
View File
@@ -3,7 +3,4 @@ set -e
sudo python3 openpilot/system/hardware/chestnut/flash.py
TARGET=openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl.chunkmanifest
rm -f "$TARGET"
SCONSFLAGS="-j4" ./openpilot/system/manager/build.py
test -s "$TARGET"
+2 -3
View File
@@ -64,9 +64,8 @@ pull_lfs() {
return
fi
# The big driving model is not used on these devices yet. Keep its pointer in
# the worktree, but don't download or copy the 1.8 GB LFS object.
LFS_EXCLUDE="openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx"
# Keep the precompiled big model as a pointer on devices without Chestnut.
LFS_EXCLUDE="openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl"
git config --local lfs.fetchexclude "$LFS_EXCLUDE"
git lfs pull --exclude="$LFS_EXCLUDE"
+5 -3
View File
@@ -49,9 +49,6 @@ for policy in /sys/devices/system/cpu/cpufreq/policy*; do
done
scons
if [ -n "$INCLUDE_BIG_MODEL" ]; then
test -f openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl.chunkmanifest
fi
if [ -z "$PANDA_DEBUG_BUILD" ]; then
# release panda fw
@@ -61,6 +58,11 @@ else
scons panda/
fi
find openpilot/selfdrive/modeld/models -name '*.pkl' -size +95M -exec ./openpilot/common/file_chunker.py {} \;
if [ -n "$INCLUDE_BIG_MODEL" ]; then
test -f openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl.chunkmanifest
fi
# Ensure no submodules in release
if test "$(git submodule--helper list | wc -l)" -gt "0"; then
echo "submodules found:"
+1 -1
View File
@@ -39,7 +39,7 @@ cd "$SOURCE_DIR"
cd "$TARGET_DIR"
rm -rf .git/modules/
find openpilot/selfdrive/modeld/models -name '*.onnx' -size +95M -exec ./openpilot/common/file_chunker.py {} \;
find openpilot/selfdrive/modeld/models -name '*.pkl' -size +95M -exec ./openpilot/common/file_chunker.py {} \;
# include source commit hash and build date in commit
GIT_HASH=$(git --git-dir="$SOURCE_DIR/.git" rev-parse HEAD)
+1 -1
View File
@@ -32,7 +32,7 @@ if __name__ == "__main__":
continue
rf = os.fsdecode(tracked_file)
if not os.getenv("INCLUDE_BIG_MODEL") and rf.startswith("openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx"):
if not os.getenv("INCLUDE_BIG_MODEL") and rf.startswith("openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl"):
continue
blacklisted = any(re.search(p, rf) for p in blacklist)
whitelisted = any(re.search(p, rf) for p in whitelist)