set var

* models: download whole-file artifacts with parallel byte ranges

HuggingFace's Xet CAS throttles each connection to roughly 1-2 MB/s but does
not limit connection count, so the single-stream _download_file left the link
idle most of the time. Replace it with a downloader that probes the size with
a one-byte Range request, creates the file at its final size and streams each
8 MB piece straight to its offset from one of 12 worker threads. The file is
never copied or assembled and only ever occupies its own size on disk. A
server that answers the probe with 200 gets one plain stream instead.
Every request sends Accept-Encoding: identity, because byte ranges count
encoded bytes while requests hands back decoded ones, so a compressed
response would land at the wrong offsets and fail as a short read.

Workers only write bytes; a reporter coroutine on the event loop publishes
progress every 0.5 s (instead of every 128 KB), re-raises the first worker
failure and polls the cancel ref. Transient errors (connection, timeout,
truncated body, HTTP 408/429/500/502/503/504) give a piece up to three
attempts; any other HTTP error fails the download at once. The whole-file
path now also publishes "verifying" while hashing before and after the
transfer, and the download is hashed against the manifest exactly as before.

A cancel has to wait for in-flight pieces to unwind (a blocked socket read
cannot be interrupted), so the row shows "canceling..." through a new
DownloadProgress.canceling flag instead of a frozen bar, and the mici cancel
button hides once the ref is gone, as it already does on the big UI.

DownloadProgress gains speed, canceling and resuming, appended to the capnp so
older readers are unaffected. All three are declared here so the wire format
lands in one place; speed and resuming are filled in by the follow-up speed
and resume changes.

Builds on #2050, which drops chunking and publishes models as whole files;
this makes that whole-file download parallel.

Measured on a comma 3X on WiFi: ~2 MB/s single stream, 17-24 MB/s steady and
30-60 MB/s peak with 12 range connections, the count picked from an on-device
sweep. Disk is not the bottleneck (eMMC writes at 56-67 MB/s with fdatasync).

Tests drive the real downloader against a local threaded range server: piece
tiling, actual parallelism (a stalled piece does not block the others), the
no-range fallback, per-piece retry, 4xx, cancel via ModelManager_DownloadRef
and mid-piece, verify/cached/downloaded state, a probe failure moving
past verifying, and hash mismatch cleanup.

Also drops two chunk-era leftovers in the tests: a stale "no chunks" note on
manifest_bundle and a one-item URL loop in the live manifest check.

Assisted-by: Claude Code (Claude Opus 4.8)
Assisted-by: Claude Code (Claude Opus 5.5)

* remove unneeded comments

* models: trim the manager tests to the regressions they catch

Each remaining test is there because a deliberate break in manager.py,
helpers.py or fetcher.py fails it. Tests that only re-checked what another
already catches are merged into a happy-path download or a table, and
checks of constants rather than behaviour are dropped. The fake server
now starts per test, keeps its state in one dataclass, and can cut a
response short.

Against 66 deliberate breaks, every one the old tests caught still fails,
and a 404 being retried, a short read and its retry, and the display
flags choosing the big model without a GPU are now caught too.

Assisted-by: Claude Code (Claude Opus 5.5)

* Update test_manager_download.py

* Update manager.py

---------

Co-authored-by: James Vecellio-Grant <159560811+Discountchubbs@users.noreply.github.com>

.github/workflows/lfs-maintenance.yaml

@@ -33,7 +33,7 @@ jobs:
 
       - name: LFS Fetch
         run: |
-          git lfs fetch
+          git lfs fetch || echo "::warning::some objects are missing on commaai/openpilot-lfs; mirroring the rest"
 
       - name: Set up Git
         run: |
@@ -68,6 +68,7 @@ jobs:
           git config lfs.activitytimeout 300
           git config lfs.transfer.maxretries 5
           git config lfs.concurrenttransfers 4
+          git config lfs.allowincompletepush true
 
       - name: Push LFS
         id: sync-and-commit
@@ -108,19 +109,19 @@ jobs:
           p = subprocess.Popen(["git", "-C", repo, "cat-file", "--batch=%(objectsize)"],
                                stdin=subprocess.PIPE, stdout=subprocess.PIPE)
           def feed():
-              for sha in small:
-                  p.stdin.write(sha.encode() + b"\n")
-              p.stdin.close()
+            for sha in small:
+              p.stdin.write(sha.encode() + b"\n")
+            p.stdin.close()
           threading.Thread(target=feed, daemon=True).start()
           rx = re.compile(rb"^version https://git-lfs\.github\.com/spec/v1\noid sha256:([0-9a-f]{64})\nsize ([0-9]+)\n")
           seen = {}
           for _ in small:
-              n = int(p.stdout.readline()); body = p.stdout.read(n); p.stdout.read(1)
-              m = rx.match(body)
-              if m:
-                  seen.setdefault(m.group(1).decode(), m.group(2).decode())
+            n = int(p.stdout.readline()); body = p.stdout.read(n); p.stdout.read(1)
+            m = rx.match(body)
+            if m:
+              seen.setdefault(m.group(1).decode(), m.group(2).decode())
           for oid, size in sorted(seen.items()):
-              print(oid, size)
+            print(oid, size)
           EOF
           for repo in . sunnypilot; do
             python3 scan_pointers.py "$repo"

.github/workflows/test_models.yml

@@ -3,9 +3,62 @@ on:
   pull_request:
     paths:
       - 'tinygrad_repo'
+      - 'openpilot/sunnypilot/modeld_v2/**'
   workflow_dispatch:
 
+env:
+  CI: 1
+  GIT_CONFIG_COUNT: 1
+  GIT_CONFIG_KEY_0: lfs.fetchexclude
+  GIT_CONFIG_VALUE_0: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
+
 jobs:
+  plain-artifact:
+    name: Plain artifact round trip and pickle converter
+    runs-on: ubuntu-24.04
+    env:
+      PYTHONPATH: ${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}
+      TINYGRAD_TESTS: 1
+    steps:
+      - uses: actions/checkout@v7
+      - run: ./tools/op.sh setup
+      - name: Build openpilot
+        run: scons -j$(nproc)
+      - name: Compile, load, cache and fuse a model with the pinned tinygrad
+        env:
+          DEV: CPU
+        run: tools/op.sh test -v openpilot/sunnypilot/modeld_v2/tests/test_plain_artifact.py
+      - name: Install LLVM 20 and ROCm comgr for the amd mock, as tinygrad_repo/.github/actions/setup-tinygrad does
+        run: |
+          . /etc/os-release
+          curl -fsSL https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc > /dev/null
+          echo "deb http://apt.llvm.org/$VERSION_CODENAME/ llvm-toolchain-$VERSION_CODENAME-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+          sudo mkdir -p /etc/apt/keyrings
+          curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+          echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $VERSION_CODENAME main" | sudo tee /etc/apt/sources.list.d/rocm.list
+          printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' | sudo tee /etc/apt/preferences.d/rocm-pin-600
+          sudo apt-get -qq update && sudo apt-get -qq install -y --no-install-recommends libllvm20 clang-20 lld-20 comgr
+          printf '/opt/rocm/lib\n/opt/rocm/lib64\n' | sudo tee /etc/ld.so.conf.d/rocm.conf && sudo ldconfig
+          echo "/usr/lib/llvm-20/bin" >> "$GITHUB_PATH"
+      - name: Convert a pickled USB+AMD jit and replay it under tinygrad's amd mock
+        env:
+          DEV: MOCKUSB+AMD
+          CHECK_OOB: 0
+          GMMU: 0
+        run: tools/op.sh test -v openpilot/sunnypilot/modeld_v2/tests/test_convert_upstream_pkl.py
+      - name: Where tinygrad moved, for the helper or failure point named above
+        if: failure() && github.event_name == 'pull_request'
+        run: |
+          git fetch -q --depth=1 origin "${{ github.base_ref }}"
+          old=$(git rev-parse "origin/${{ github.base_ref }}:tinygrad_repo") new=$(git rev-parse HEAD:tinygrad_repo)
+          git -C tinygrad_repo fetch -q origin "$old" || true
+          echo "tinygrad_repo $old -> $new, in what helpers.py and plain_artifact_compiler.py read:"
+          git -C tinygrad_repo diff "$old" "$new" -- tinygrad/uop/ops.py tinygrad/engine/jit.py tinygrad/engine/realize.py tinygrad/device.py \
+            tinygrad/helpers.py tinygrad/dtype.py tinygrad/renderer/__init__.py
+          echo "and in what convert_upstream_pkl.py decodes:"
+          git -C tinygrad_repo diff --stat "$old" "$new" -- tinygrad/runtime/autogen/am/pm4_nv.py tinygrad/runtime/autogen/amd_gpu.py
+          git -C tinygrad_repo diff "$old" "$new" -- tinygrad/runtime/support/hcq2.py tinygrad/runtime/support/usb.py tinygrad/runtime/ops_amd.py
+
   generate-matrix:
     runs-on: ubuntu-latest
     outputs:

openpilot/cereal/custom.capnp

@@ -138,6 +138,9 @@ struct ModelManagerSP @0xaedffd8f31e7b55d {
     status @0 :DownloadStatus;
     progress @1 :Float32;
     eta @2 :UInt32;
+    speed @3 :Float32;
+    canceling @4 :Bool;
+    resuming @5 :Bool;
   }
 
   struct Chunk {

openpilot/selfdrive/modeld/modeld.py

@@ -9,7 +9,6 @@ os.environ['GMMU'] = '0' # for chestnut fast loading, noop for qcom
 os.environ.setdefault('AM_POWER_LIMIT', '100')
 os.environ.setdefault('AM_PERF_LEVEL', 'high')
 from tinygrad.device import Buffer, Device
-from tinygrad.engine.realize import lower_and_compile
 from tinygrad.dtype import DType, dtypes
 from tinygrad.engine.realize import lower_and_compile
 from tinygrad.tensor import Tensor

openpilot/selfdrive/navigation/helpers.py

@@ -1,7 +1,4 @@
 import time
-from typing import Any
-
-import numpy as np
 
 
 def latest_gps_source(sm) -> str | None:
@@ -10,34 +7,3 @@ def latest_gps_source(sm) -> str | None:
   sources = [s for s in ('gpsLocation', 'gpsLocationExternal')
              if sm.valid[s] and sm[s].hasFix and 0 <= now - sm.recv_time[s] < 2]
   return max(sources, key=lambda s: sm.logMonoTime[s]) if sources else None
-
-
-# convert mapbox steps to valhalla enums the model was trained on
-VALHALLA_COMMAND_COUNT = 46
-VALHALLA_SIDED_TYPES = {  # (straight, right, left)
-  'depart': (1, 2, 3), 'arrive': (4, 5, 6),
-  'on ramp': (17, 18, 19), 'off ramp': (20, 20, 21), 'fork': (22, 23, 24),
-  'roundabout': (26, 26, 26), 'rotary': (26, 26, 26), 'roundabout turn': (26, 26, 26),
-  'exit roundabout': (27, 27, 27), 'exit rotary': (27, 27, 27), 'notification': (28, 28, 28),
-}
-VALHALLA_TURN_TYPES = {'straight': 8, 'slight right': 9, 'right': 10, 'sharp right': 11, 'uturn': 13,
-                       'sharp left': 14, 'left': 15, 'slight left': 16}  # turn, continue, new name, end of road
-
-
-def valhalla_type(maneuver_type: str, modifier: str) -> int | None:
-  if maneuver_type == 'merge':
-    return None  # Valhalla gives no maneuver at a merge, so the model never saw one
-  if maneuver_type in VALHALLA_SIDED_TYPES:
-    straight, right, left = VALHALLA_SIDED_TYPES[maneuver_type]
-    return right if 'right' in modifier else left if 'left' in modifier else straight
-  return VALHALLA_TURN_TYPES.get(modifier, 8)
-
-
-def navigation_input(maneuvers: Any, traveled: float) -> np.ndarray:
-  navigation = np.zeros(VALHALLA_COMMAND_COUNT + 1, dtype=np.float32)
-  for maneuver in maneuvers:
-    command, distance = valhalla_type(maneuver.type, maneuver.modifier), maneuver.distance - traveled
-    if command is not None and command > 6 and distance >= 0:
-      navigation[command], navigation[-1] = 1, distance / 100
-      break
-  return navigation

openpilot/selfdrive/ui/sunnypilot/layouts/settings/models.py

@@ -235,6 +235,8 @@ class ModelsLayout(Widget):
     progress = sum(p.progress for p in progresses) / len(progresses)
     ds = custom.ModelManagerSP.DownloadStatus
 
+    if any(p.canceling for p in progresses):
+      return {"name": name, "downloading": True, "progress": progress, "status_text": tr("canceling...")}
     if ds.failed in statuses:
       # close.png is authored black and a tint cannot lift it, hence close2
       return {"name": name, "status_text": tr("download failed"), "text_color": rl.RED, "icon": "icons/close2.png"}

openpilot/selfdrive/ui/sunnypilot/mici/layouts/models.py

@@ -251,11 +251,12 @@ class ModelsLayoutMici(NavScroller):
       self.current_model_info.info_text.set_text(tr("download failed"))
 
     elif manager.selectedBundle and manager.selectedBundle.status == custom.ModelManagerSP.DownloadStatus.downloading:
-      self.cancel_download_btn.set_visible(True)
+      self.cancel_download_btn.set_visible(ui_state.params.get("ModelManager_DownloadRef") is not None)
       device.set_override_interactive_timeout(5)
       progress = 0.0
       count = 0
       verifying = False
+      canceling = False
       for model in manager.selectedBundle.models:
         count += 1
         p = model.artifact.downloadProgress
@@ -263,11 +264,13 @@ class ModelsLayoutMici(NavScroller):
                         custom.ModelManagerSP.DownloadStatus.verifying):
           progress += p.progress
           verifying = verifying or p.status == custom.ModelManagerSP.DownloadStatus.verifying
+          canceling = canceling or p.canceling
         elif p.status in (custom.ModelManagerSP.DownloadStatus.downloaded,
                           custom.ModelManagerSP.DownloadStatus.cached):
           progress += 100.0
 
-      self.current_model_info.current_model_header.set_text(tr("verifying") if verifying else tr("downloading"))
+      header = tr("canceling...") if canceling else tr("verifying") if verifying else tr("downloading")
+      self.current_model_info.current_model_header.set_text(header)
       self.cancel_download_btn.set_text(tr("cancel verification") if verifying else tr("cancel download"))
       self.current_model_info.current_model_header._shimmer = True
       name_text = manager.selectedBundle.internalName.lower()

openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py

@@ -7,6 +7,7 @@ See the LICENSE.md file in the root directory for more details.
 """
 import argparse
 import dataclasses
+import inspect
 import itertools
 import mmap
 import re
@@ -45,6 +46,11 @@ def _address(v):
   return _root(v.src[0]), add + start, size
 
 
+def _buffer(u):
+  from tinygrad.uop.ops import Ops
+  return u.buffer if u.op is Ops.BUFFER else getattr(u.arg, "buffer", None)
+
+
 def _patched(buf, topo) -> tuple[bytearray, dict]:
   from tinygrad.uop.ops import Ops
   blob, rows = bytearray(buf.max_numel() * buf.dtype.itemsize), {}
@@ -72,7 +78,7 @@ def _evaluate(u, env):
     table, index = _root(u.src[0].src[0]), _evaluate(u.src[0].src[1], env)
     return struct.unpack_from(u.dtype.fmt, table.arg, index * u.dtype.itemsize)[0] if table.op is Ops.BINARY and index is not None else None
   if u.op in GroupOp.ALU and None not in (values := [_evaluate(x, env) for x in u.src]):
-    return exec_alu(u.op, u.dtype, values, False)
+    return exec_alu(u.op, u.dtype, values, *([False] if "truncate_output" in inspect.signature(exec_alu).parameters else []))
   return None
 
 
@@ -112,7 +118,7 @@ def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
   from tinygrad.uop.ops import Ops
   submit = jit.captured._linear.src[0].without_after
   kernels, topo = submit.arg.aux.kernels, list(jit.captured._linear.toposort(enter_calls=True))
-  allocs = {str(u.tag): u for u in topo if u.op is Ops.ALLOC and u.tag is not None}
+  allocs = {str(name): u for u in topo if u.op is Ops.ALLOC and (name := getattr(u, "tag", None) or u.arg.name) is not None}
   (commands, patched), kernargs = _patched(allocs["cmdbuf_compute_0"], topo), allocs["kernargs_compute_0"]
   argument_words, pokes = _patched(kernargs, topo)[1], _input_pokes(submit, kernargs)
   sh = {name: getattr(amd_gpu, f"reg{name}") + amd_gpu.GC_BASE__INST0_SEG0 - pm4.PACKET3_SET_SH_REG_START
@@ -158,7 +164,7 @@ def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
         args.append({"input": pokes[slot][0]})
       elif slot in argument_words:
         base, at, nbytes = _address(argument_words[slot])
-        args.append({"buffer": buffers.setdefault(base, len(buffers)), "offset": at, "nbytes": nbytes or base.arg.buffer.nbytes, "dtype": u8})
+        args.append({"buffer": buffers.setdefault(base, len(buffers)), "offset": at, "nbytes": nbytes or base.max_numel() * base.dtype.itemsize, "dtype": u8})
     outs, ins = kernel[6]
     calls.append({"kind": "program", "name": re.sub(r"\x1b\[[0-9;]*m", "", kernel[1]), "lib": programs[image_alloc], "target": dataclasses.asdict(target),
                   "signature": [[f"data{i}", i, u8, [a.get("nbytes", 1)]] for i, a in enumerate(args)], "global_size": grid, "local_size": local,
@@ -166,12 +172,13 @@ def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
     needs_init |= {args[i]["buffer"] for i in ins if "buffer" in args[i] and args[i]["buffer"] not in written}
     written |= {args[i]["buffer"] for i in outs if "buffer" in args[i]}
 
+  assert all(_buffer(b) is not None for b, i in buffers.items() if i in needs_init), "a kernel reads a scratch before any kernel writes it"
   names = list(jit.captured.expected_names)
   inputs = [{"slot": slot, "name": names[slot] if slot < len(names) else None, "dtype": repr(u.dtype), "numel": int(u.max_numel()), "device": u.device}
             for slot, u in enumerate(_prepare_jit_inputs((), kwargs)[0])]
   header = {"input_names": names, "calls": calls, "returns": {"kind": "NoneType", "tensors": []}, "inputs": inputs,
-            "buffers": [{"nbytes": b.arg.buffer.nbytes, "device": b.arg.buffer.device, "dtype": u8}
-                        | ({"init": blob.add(b.arg.buffer.ensure_allocated().as_memoryview())} if i in needs_init else {}) for b, i in buffers.items()],
+            "buffers": [{"nbytes": bb.nbytes if (bb := _buffer(b)) else b.max_numel() * b.dtype.itemsize, "device": bb.device if bb else b.device, "dtype": u8}
+                        | ({"init": blob.add(bb.ensure_allocated().as_memoryview())} if i in needs_init else {}) for b, i in buffers.items()],
             "check": {"inputs": [[path, store(array), device] for path, array, device in frame["inputs"]],
                       "outputs": [store(array) for array in frame["outputs"]], "output_container": frame["output_container"]}}
   with sink:

openpilot/sunnypilot/modeld_v2/frame_fusion.py

@@ -9,16 +9,14 @@ import itertools
 import math
 from functools import partial
 
-import numpy as np
-
 from openpilot.common.basedir import BASEDIR
 from openpilot.sunnypilot.modeld_v2.helpers import _dtype, read_parts
 
 
-def pack_layout(specs: dict[str, tuple], frame_size: int) -> tuple[dict[str, int], int]:
+def pack_layout(shapes: dict[str, tuple], frame_size: int) -> tuple[dict[str, int], int]:
   offsets, offset = {}, 0
-  for name, (shape, dtype) in ({"tfm": ((2, 3, 3), np.float32)} | specs).items():
-    offsets[name], offset = offset, offset + -(-math.prod(shape) * np.dtype(dtype).itemsize // 128) * 128
+  for name, shape in ({"tfm": (2, 3, 3)} | shapes).items():
+    offsets[name], offset = offset, offset + -(-math.prod(shape) * 4 // 128) * 128
   return offsets | {"frames": offset}, offset + 2 * frame_size
 
 
@@ -40,7 +38,7 @@ def frame_part(warp_path: str, parts: dict, extras: dict) -> dict:
     return {}
   (model, model_blob), (warp, warp_blob) = parts["run"], read_parts(warp_path)[0]["run"]
   states = sorted(name for name in extras["input_specs"] if f"next_{name}" in extras["metadata"]["output_shapes"])
-  host = {name: (shape, dtype) for name, (shape, dtype, _) in extras["input_specs"].items() if name not in states and name != "new_img"}
+  host = {name: shape for name, (shape, _, _) in extras["input_specs"].items() if name not in states and name != "new_img"}
   frame_size = next(s["numel"] for s in warp["inputs"] if s["name"] == "input_frame") // 2
   offsets, packed = pack_layout(host, frame_size)
   shift, upload, u8 = len(model_blob), len(model["buffers"]) + len(warp["buffers"]), repr(_dtype("dtypes.uint8"))

openpilot/sunnypilot/modeld_v2/helpers.py

@@ -250,6 +250,12 @@ def _allocate_buffers(header: dict, blob: memoryview, device_map: dict | None, s
   return buffers
 
 
+def _fields(what, /, **values):  # a tinygrad dataclass built or replaced with the fields it has today: a dropped field is fine, a new required one is not
+  known = {field.name for field in dataclasses.fields(what)}
+  values = {name: value for name, value in values.items() if name in known}
+  return what(**values) if isinstance(what, type) else dataclasses.replace(what, **values)
+
+
 def _program_uop(call: dict, lib: bytes):
   from tinygrad.helpers import Target
   from tinygrad.uop.ops import KernelInfo, Ops, ProgramInfo, UOp
@@ -257,9 +263,9 @@ def _program_uop(call: dict, lib: bytes):
   for name, compact_slot, dtype, shape in call["signature"]:
     parameter = UOp.param(call["globals"][compact_slot], _dtype(dtype), math.prod(shape))
     image = (shape[0], shape[1]) if len(shape) == 3 and shape[2] == 4 else None
-    parameters.append(parameter.replace(arg=dataclasses.replace(parameter.arg, name=name, image=image)))
-  info = ProgramInfo(global_size=tuple(call["global_size"]), local_size=tuple(call["local_size"]), vars=(), globals=tuple(call["globals"]),
-                     outs=tuple(call["outs"]), ins=tuple(call["ins"]), target=Target(**call["target"]))
+    parameters.append(parameter.replace(arg=_fields(parameter.arg, name=name, image=image)))
+  info = _fields(ProgramInfo, global_size=tuple(call["global_size"]), local_size=tuple(call["local_size"]), globals=tuple(call["globals"]),
+                 outs=tuple(call["outs"]), ins=tuple(call["ins"]), target=_fields(Target, **call["target"]))
   return UOp(Ops.PROGRAM, src=(UOp(Ops.SINK, arg=KernelInfo(name=call["name"])), UOp(Ops.LINEAR, src=tuple(parameters)),
                                UOp(Ops.SOURCE, arg=""), UOp(Ops.BINARY, arg=lib)), arg=info)
 

openpilot/sunnypilot/modeld_v2/model_adapters.py

@@ -172,12 +172,12 @@ class NativeTinygradAdapter(BaseModelAdapter):
 
     self.input_queues = {name: Tensor(np.zeros(shape, dtype=dtype), device=self.model_device).realize()
                          for name, (shape, dtype) in self.input_shapes.items() if name in self.state_pairs}
-    specs = {name: spec for name, spec in self.input_shapes.items() if name not in self.state_pairs and name != 'new_img'}
-    offsets, packed = pack_layout(specs, self.warp_frame_size)
+    shapes = {name: shape for name, (shape, _) in self.input_shapes.items() if name not in self.state_pairs and name != 'new_img'}
+    offsets, packed = pack_layout(shapes, self.warp_frame_size)
     self.packed_input = np.zeros(packed, dtype=np.uint8)
     self.host = Tensor(self.packed_input, device='NPY')
-    self.numpy_inputs = {name: np.ndarray(shape, dtype=dtype, buffer=self.packed_input, offset=offsets[name])
-                         for name, (shape, dtype) in ({'tfm': ((2, 3, 3), np.dtype(np.float32))} | specs).items()}
+    self.numpy_inputs = {name: np.ndarray(shape, dtype=np.float32, buffer=self.packed_input, offset=offsets[name])
+                         for name, shape in ({'tfm': (2, 3, 3)} | shapes).items()}
     self.frames = self.packed_input[offsets['frames']:].reshape(2, self.warp_frame_size)
 
     self.outputs = {name: Tensor(np.zeros(shape, dtype=dtype), device=device).realize() for name, (shape, dtype, device) in self.jits['output_specs'].items()}

openpilot/sunnypilot/modeld_v2/modeld.py

@@ -35,7 +35,6 @@ from openpilot.common.transformations.camera import DEVICE_CAMERAS
 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
-from openpilot.selfdrive.navigation.helpers import navigation_input
 from openpilot.selfdrive.modeld.modeld import ChestnutGpuState
 
 from openpilot.sunnypilot.modeld_v2.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState, get_curvature_from_output
@@ -173,9 +172,6 @@ class ModelState(ModelStateBase):
     for key in ('traffic_convention', 'lateral_control_params', 'action_t'):
       if key in self.numpy_inputs and key in inputs:
         self.numpy_inputs[key][:] = inputs[key]
-    for name in ('nav_map', 'navigation'):
-      if name in self.numpy_inputs:
-        self.numpy_inputs[name][:] = inputs.get(name, 0)
 
     if self.adapter.is_native:
       for i, key in enumerate(self._vision_input_names):
@@ -257,10 +253,6 @@ class ModelState(ModelStateBase):
     return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature), desiredAcceleration=float(desired_accel), shouldStop=bool(stop))
 
 
-def takes_nav(*models) -> bool:
-  return any(name in m.numpy_inputs for m in models if m is not None for name in ('nav_map', 'navigation'))
-
-
 def main(demo=False):
   cloudlog.warning("modeld init")
 
@@ -290,7 +282,6 @@ def main(demo=False):
   vipc_client_main_stream = VisionStreamType.VISION_STREAM_WIDE_ROAD if main_wide_camera else VisionStreamType.VISION_STREAM_NARROW_ROAD
   vipc_client_main = VisionIpcClient("camerad", vipc_client_main_stream, True)
   vipc_client_extra = VisionIpcClient("camerad", VisionStreamType.VISION_STREAM_WIDE_ROAD, False)
-  vipc_client_map = VisionIpcClient("mapsd", VisionStreamType.VISION_STREAM_MAP, True)
   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):
@@ -341,9 +332,8 @@ def main(demo=False):
   # messaging
   pub_socks = ["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"] + (["chestnutGpuState"] if CHESTNUT else [])
   pm = PubMaster(pub_socks)
-  nav = takes_nav(model, small_model)
   sm = SubMaster(["deviceState", "carState", "narrowRoadCameraState", "extrinsicsCalibration", "driverMonitoringState",
-                  "carControl", "controlsState", "lateralDelay"] + (["navInstruction"] if nav else []))
+                  "carControl", "controlsState", "lateralDelay"])
 
   publish_state = PublishState()
   chestnut_state = ChestnutGpuState(pm, model.chestnut) if CHESTNUT else None
@@ -376,7 +366,6 @@ def main(demo=False):
   DH = DesireHelper()
   meta_constants = load_meta_constants()
   RELC = RoadEdgeLaneChangeController()
-  nav_map, nav_map_time = None, 0.
 
   while True:
     # Keep receiving frames until we are at least 1 frame ahead of previous extra frame
@@ -466,16 +455,6 @@ def main(demo=False):
     if 'action_t' in model.numpy_inputs:
       inputs['action_t'] = np.array([lat_action_t, long_action_t], dtype=np.float32)
 
-    now = time.monotonic()
-    if nav and (vipc_client_map.is_connected() or vipc_client_map.connect(False)):
-      if (buf_map := vipc_client_map.recv(0)) is not None:
-        nav_map = np.frombuffer(buf_map.data, dtype=np.uint8, count=buf_map.width * buf_map.height).reshape(buf_map.height, buf_map.width).copy()
-        nav_map_time = now
-    if nav_map is not None and now - nav_map_time < 0.5: # map timeout
-      inputs['nav_map'] = nav_map
-    if nav and sm.valid['navInstruction'] and now - sm.recv_time['navInstruction'] < 2:
-      inputs['navigation'] = navigation_input(sm['navInstruction'].allManeuvers, v_ego * (now - sm.recv_time['navInstruction']))
-
     mt1 = time.perf_counter()
     try:
       send_chestnut = (chestnut_state is not None and

openpilot/sunnypilot/modeld_v2/plain_artifact_compiler.py

@@ -23,7 +23,7 @@ def record_kernel_linear():
   recorded, hcq_compile, compile_linear = [], realize.hcq_compile, jit.compile_linear
 
   def recording_compile_linear(*args, **kwargs):
-    with mock.patch.object(realize, "hcq_compile", lambda linear, *a, **k: (recorded.append(linear), hcq_compile(linear, *a, **k))[1]):
+    with mock.patch.object(realize, "hcq_compile", lambda linear, *a, **k: (linear.src and recorded.append(linear), hcq_compile(linear, *a, **k))[1]):
       return compile_linear(*args, **kwargs)
 
   with mock.patch.object(jit, "compile_linear", recording_compile_linear):
@@ -38,6 +38,18 @@ def capture_frame(jit, kwargs: dict, output_container: str | None = None):
   return {"inputs": inputs, "outputs": [tensor.numpy().copy() for tensor in outputs], "output_container": output_container}, result
 
 
+def kernel_signature(ast, elf) -> list:
+  from tinygrad.dtype import AddrSpace
+  if hasattr(ast, "kernel_params"):
+    if any(p.arg.addrspace is AddrSpace.ALU for p in ast.kernel_params):
+      raise ValueError(f"{elf.name} takes a variable, which plain artifacts don't support")
+    compact = {slot: i for i, slot in enumerate(ast.arg.globals)}
+    return [[p.arg.name, compact[p.arg.slot], repr(p.dtype), [int(x) for x in p._shape]] for p in ast.kernel_params]
+  if not all(isinstance(slot, int) for _, slot, _, _ in elf.signature):
+    raise ValueError("TinyELF.signature lost its slots and this tinygrad has no UOp.kernel_params: update kernel_signature")
+  return [[name, slot, repr(dtype), [int(x) for x in shape]] for name, slot, dtype, shape in elf.signature]
+
+
 def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict, memoryview]:
   from tinygrad.uop.ops import Ops
   sink = tempfile.TemporaryFile(dir=".")
@@ -51,6 +63,12 @@ def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict,
     if uop.op is Ops.PARAM:
       input_specs.setdefault(uop.arg.slot, {"dtype": repr(uop.dtype), "numel": int(uop.max_numel()), "device": uop.device})
       return {"input": uop.arg.slot}
+    root, offset = uop, 0
+    while root.op in (Ops.AFTER, Ops.BITCAST, Ops.SHRINK):
+      offset, root = offset + (root.src[1].arg * root.src[0].dtype.itemsize if root.op is Ops.SHRINK else 0), root.src[0]
+    if root.op is Ops.ALLOC:
+      base = bases.setdefault(id(root), (len(bases), root, root.dtype))
+      return {"buffer": base[0], "offset": offset, "nbytes": int(uop.max_numel()) * uop.dtype.itemsize, "dtype": repr(uop.dtype)}
     buffer = uop.buffer
     root = buffer._base if buffer._base is not None else buffer
     base = bases.setdefault(id(root), (len(bases), root, uop.base.dtype if buffer._base is None else uop.dtype))
@@ -63,7 +81,7 @@ def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict,
       elf, info = ast.to_elf(), ast.arg
       global_size, local_size = info.launch_dims({})
       calls.append({"kind": "program", "name": elf.name, "lib": blob.add(bytes(elf.lib)), "target": dataclasses.asdict(elf.target),
-                    "signature": [[name, slot, repr(dtype), [int(x) for x in shape]] for name, slot, dtype, shape in elf.signature],
+                    "signature": kernel_signature(ast, elf),
                     "global_size": [int(x) for x in global_size], "local_size": [int(x) for x in local_size],
                     "globals": list(info.globals), "outs": list(info.outs), "ins": list(info.ins), "args": arguments})
       outs, ins = info.outs, info.ins
@@ -77,7 +95,9 @@ def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict,
 
   ret = jit.captured.ret
   returns = {"kind": type(ret).__name__, "tensors": [describe(t.uop.base) | {"shape": [int(x) for x in t.shape]} for t in flatten(ret)]}
-  buffers = [{"nbytes": base.nbytes, "device": base.device, "dtype": repr(dtype)}
+  from tinygrad.device import Buffer
+  assert not [base for index, base, _ in bases.values() if index in needs_init and not isinstance(base, Buffer)], "a kernel reads a scratch before writing it"
+  buffers = [{"nbytes": base.nbytes if isinstance(base, Buffer) else int(base.max_numel()) * dtype.itemsize, "device": base.device, "dtype": repr(dtype)}
              | ({"init": blob.add(base.as_memoryview())} if index in needs_init else {}) for index, base, dtype in bases.values()]
   names = list(jit.captured.expected_names)
   if check is not None:

openpilot/sunnypilot/modeld_v2/tests/test_convert_upstream_pkl.py

@@ -0,0 +1,121 @@
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import os
+import re
+import sys
+import tempfile
+import unittest
+from unittest import mock
+import numpy as np
+
+os.environ.setdefault("CHECK_OOB", "0")
+os.environ.setdefault("GMMU", "0")
+if sys.platform == "darwin" and os.environ.get("DEV", "").startswith("MOCKUSB"):  # the mock's VRAM is a memfd, which macOS lacks: an unlinked file is the same
+  from tinygrad.runtime.autogen import libc
+
+  def _memfd_create(name, flags):
+    fd, path = tempfile.mkstemp()
+    os.unlink(path)
+    return fd
+  libc.memfd_create = _memfd_create
+
+from openpilot.common.test import OpenpilotTestCase
+from openpilot.sunnypilot.modeld_v2 import convert_upstream_pkl
+from openpilot.sunnypilot.modeld_v2.convert_upstream_pkl import convert
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, load_model
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame
+
+IMG, DESIRE, STATE, OUT = (6, 8, 8), (8,), (4, 16), (1, 32)
+
+
+class TestConvertUpstreamPkl(OpenpilotTestCase):
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    if not os.environ.get("TINYGRAD_TESTS"):
+      raise unittest.SkipTest("TINYGRAD_TESTS not set. skipping converter test")
+    from tinygrad import Device, Tensor
+    assert Device.DEFAULT == "AMD" and getattr(Device["AMD"], "is_usb", False), f"{Device.DEFAULT} is not an AMD device behind the USB bridge"
+    assert (Tensor([1, 2, 3]) * 2).tolist() == [2, 4, 6]
+    cls.tmp = tempfile.TemporaryDirectory()
+    cls.rng = np.random.default_rng(0)
+    cls.weights = Tensor(cls.rng.standard_normal((16, 32)).astype(np.float32)).realize()
+    cls.bias = Tensor(cls.rng.standard_normal((1, 32)).astype(np.float32)).realize()
+    cls.input_specs = {"input_imgs": (IMG, "uint8", Device.DEFAULT), "desire": (DESIRE, "float32", Device.DEFAULT),
+                       "features": (STATE, "float32", Device.DEFAULT)}
+    cls.output_specs = {"outputs": (OUT, "float32", Device.DEFAULT), "next_features": (STATE, "float32", Device.DEFAULT)}
+
+  @classmethod
+  def tearDownClass(cls):
+    cls.tmp.cleanup()
+    super().tearDownClass()
+
+  def model(self, input_imgs, desire, features, output_buffers):
+    from tinygrad import Tensor
+    feat = features.mean(axis=0).reshape(1, 16) + input_imgs.float().mean() + desire.sum()
+    Tensor.realize(output_buffers["outputs"].assign((feat @ self.weights + self.bias).relu()),
+                   output_buffers["next_features"].assign(features[1:].cat(feat * 2)))
+
+  def inputs(self, seed: int) -> dict:
+    from examples.openpilot.helpers import allocate_inputs
+    rng = np.random.default_rng(seed)
+
+    def initialize(arrays: dict):
+      for value in arrays.values():
+        value[...] = rng.standard_normal(value.shape) if value.dtype.kind == "f" else rng.integers(0, 256, value.shape)
+    return allocate_inputs(self.input_specs, initialize) | {"output_buffers": allocate_inputs(self.output_specs)}
+
+  def jit(self):
+    from tinygrad import TinyJit
+    run = TinyJit(self.model, prune=True)
+    kwargs = self.inputs(1)
+    run(**kwargs)
+    run(**kwargs)
+    return run
+
+  def assert_matches_eager(self, plain, seed: int):
+    fresh, expected = self.inputs(seed), self.inputs(seed)
+    plain(**fresh)
+    self.model(**expected)
+    for name in self.output_specs:
+      assert np.array_equal(fresh["output_buffers"][name].numpy(), expected["output_buffers"][name].numpy()), name
+
+  def convert(self, jit) -> tuple[dict, memoryview]:
+    from tinygrad import Device
+    kwargs = self.inputs(2)
+    frame, _ = capture_frame(jit, kwargs, "output_buffers")
+    header, blob = convert(jit, kwargs, frame, Device[Device.DEFAULT].renderer.target)
+    kernels = jit.captured._linear.src[0].without_after.arg.aux.kernels
+    assert [call["name"] for call in header["calls"]] == [re.sub(r"\x1b\[[0-9;]*m", "", kernel[1]) for kernel in kernels]
+    assert all(re.fullmatch(r"\w+", call["name"]) for call in header["calls"]), "kernel names: failure point 1"
+    assert [(tuple(call["outs"]), tuple(call["ins"])) for call in header["calls"]] == [tuple(kernel[6]) for kernel in kernels], "outs/ins: failure point 1"
+    bound = {argument["input"] for call in header["calls"] for argument in call["args"] if "input" in argument}
+    assert bound == {spec["slot"] for spec in header["inputs"]}, f"input slots bound by pokes {sorted(bound)}: failure point 4"
+    assert [spec["name"] for spec in header["inputs"]][:len(self.input_specs)] == sorted(self.input_specs)
+    return header, blob
+
+  def test_converted_jit_matches_eager(self):  # the submit decoded into plain calls, replayed against its own frame
+    jit = self.jit()
+    plain = PlainJit(*self.convert(jit))
+    self.assert_matches_eager(plain, 3)
+    self.assert_matches_eager(plain, 4)
+
+  def test_pickle_round_trip_through_main(self):
+    from examples.openpilot.helpers import dump_pickle, load_pickle
+    from tinygrad.engine.realize import lower_and_compile
+    pkl, out = os.path.join(self.tmp.name, "tiny_tinygrad.pkl"), os.path.join(self.tmp.name, "tiny_plain.pkl")
+    dump_pickle({"run": self.jit(), "input_specs": self.input_specs, "output_specs": self.output_specs, "metadata": {"name": "tiny"}}, pkl)
+
+    artifact = load_pickle(pkl, out_of_band=True)
+    artifact["run"].captured._linear = lower_and_compile(artifact["run"].captured._linear)
+    self.assert_matches_eager(PlainJit(*self.convert(artifact["run"])), 5)
+
+    with mock.patch.object(sys, "argv", ["convert_upstream_pkl", pkl, out]):
+      convert_upstream_pkl.main()
+    bundle = load_model(out, cache=False, verify=True)
+    assert bundle["metadata"] == {"name": "tiny"} and bundle["input_specs"] == self.input_specs and bundle["output_specs"] == self.output_specs
+    self.assert_matches_eager(bundle["run"], 6)

openpilot/sunnypilot/modeld_v2/tests/test_plain_artifact.py

@@ -0,0 +1,162 @@
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import dataclasses
+import functools
+import inspect
+import os
+import sys
+import tempfile
+from unittest import mock
+import numpy as np
+import unittest
+
+import openpilot.sunnypilot.modeld_v2.helpers as helpers
+from openpilot.common.test import OpenpilotTestCase
+from openpilot.sunnypilot.modeld_v2 import compile_warp
+from openpilot.sunnypilot.modeld_v2.frame_fusion import frame_part, pack_layout
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, cache_is_fresh, load_model, read_parts, write_bundle
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame, record_kernel_linear, serialize_jit
+
+CAM_W, CAM_H, MODEL_W, MODEL_H = 32, 16, 16, 8
+FRAME = f"{CAM_W},{CAM_H},{CAM_W},{CAM_H},{CAM_H // 2},{CAM_W * CAM_H * 3 // 2}"
+FRAME_SIZE = CAM_W * CAM_H * 3 // 2
+HOST = {"desire": (8,), "traffic_convention": (2,)}
+STATE, OUT = (4, 16), (1, 32)
+
+
+def compile_jit(fn, inputs: dict, output_container: str | None = None) -> tuple[dict, memoryview]:
+  from tinygrad import TinyJit
+  run = TinyJit(fn, prune=True)
+  with record_kernel_linear() as recorded:
+    run(**inputs)
+    run(**inputs)
+    check, _ = capture_frame(run, inputs, output_container)
+  assert len(recorded) == 1, f"expected one captured kernel list, got {len(recorded)}"
+  return serialize_jit(run, recorded[0], check)
+
+
+class TestPlainArtifact(OpenpilotTestCase):
+  @classmethod
+  def setUpClass(cls):
+    super().setUpClass()
+    if not os.environ.get("TINYGRAD_TESTS"):
+      raise unittest.SkipTest("TINYGRAD_TESTS env var not set, skipping artifact test.")
+    from tinygrad import Tensor
+    cls.tmp = tempfile.TemporaryDirectory()
+    cls.rng = np.random.default_rng(0)
+    cls.weights = Tensor(cls.rng.standard_normal((16, 32)).astype(np.float32)).realize()
+    cls.bias = Tensor(cls.rng.standard_normal((1, 32)).astype(np.float32)).realize()
+
+  @classmethod
+  def tearDownClass(cls):
+    cls.tmp.cleanup()
+    super().tearDownClass()
+
+  def model(self, x, h):  # a two input, one weight model: the shape of a policy without the size
+    return ((x @ self.weights + self.bias).relu() + h.mean()).contiguous()
+
+  def tensors(self, seed: int):
+    from tinygrad import Tensor
+    rng = np.random.default_rng(seed)
+    return {"x": Tensor(rng.standard_normal((1, 16)).astype(np.float32)).realize(), "h": Tensor(rng.standard_normal((4, 16)).astype(np.float32)).realize()}
+
+  def artifact(self, name: str = "model") -> str:
+    path = os.path.join(self.tmp.name, f"{name}_tinygrad.pkl")
+    if not os.path.exists(path):
+      write_bundle(path, {"run": compile_jit(self.model, self.tensors(1))}, {"metadata": {"name": name}})
+    return path
+
+  def test_round_trip_matches_jit(self):  # compiler -> bundle -> PlainJit, replay verified, then fresh inputs bit exact with eager tinygrad
+    bundle = load_model(self.artifact(), cache=False, verify=True)
+    assert bundle["metadata"] == {"name": "model"} and isinstance(bundle["run"], PlainJit)
+    fresh = self.tensors(2)
+    assert np.array_equal(bundle["run"](**fresh).numpy(), self.model(**fresh).numpy())
+
+  def test_replay_check_catches_a_changed_kernel(self):
+    parts, extras = read_parts(self.artifact())
+    header, blob = parts["run"]
+    poisoned = dict(header, check=dict(header["check"], outputs=[header["check"]["inputs"][0][1]]))
+    with self.assertRaises(helpers.ArtifactMismatch):
+      PlainJit(poisoned, blob)
+
+  def test_cache_round_trip_holds_no_weights_and_is_rebuilt_when_stale(self):
+    path = self.artifact("cached")
+    expected = load_model(path, cache=False, verify=False)["run"](**self.tensors(3)).numpy()
+    load_model(path)
+    assert cache_is_fresh(path)
+    with open(f"{path}.cache", "rb") as file:
+      cache = file.read()
+    assert self.weights.numpy().tobytes() not in cache and len(cache) < os.path.getsize(path)
+    with mock.patch.object(helpers, "PlainJit", side_effect=AssertionError("a fresh cache must not rebuild")):
+      cached = load_model(path)
+    assert np.array_equal(cached["run"](**self.tensors(3)).numpy(), expected)
+    with mock.patch.object(helpers, "tinygrad_identity", return_value="another tinygrad"):
+      assert not cache_is_fresh(path)
+      assert isinstance(load_model(path)["run"], PlainJit)
+      assert cache_is_fresh(path)
+    assert not cache_is_fresh(path)  # keyed on the tinygrad that wrote it, so the next boot after a bump rebuilds once
+    os.utime(path)
+    with mock.patch.object(helpers, "tinygrad_identity", return_value="another tinygrad"):
+      assert not cache_is_fresh(path)
+
+  def test_fused_with_warp_matches_warp_then_model(self):
+    from tinygrad import Tensor
+    warp_path = os.path.join(self.tmp.name, "warp_tinygrad.pkl")
+    with mock.patch.object(sys, "argv", ["compile_warp", "--frame", FRAME, "--warp-to", f"{MODEL_W}x{MODEL_H}", "--output", warp_path]):
+      compile_warp.main()
+    warp = load_model(warp_path, cache=False)["run"]
+
+    def native(new_img, desire, traffic_convention, state_feat_q, output_buffers):
+      feat = state_feat_q.mean(axis=0).reshape(1, 16) + new_img.float().mean() + desire.sum() + traffic_convention[0]
+      Tensor.realize(output_buffers["outputs"].assign((feat @ self.weights + self.bias).relu()),
+                     output_buffers["next_state_feat_q"].assign(state_feat_q[1:].cat(feat * 2)))
+
+    def model_inputs(new_img):
+      return {"new_img": new_img} | {name: Tensor(np.ones(shape, np.float32) * 0.5).realize() for name, shape in HOST.items()} | \
+             {"state_feat_q": Tensor(self.rng.standard_normal(STATE).astype(np.float32)).realize(),
+              "output_buffers": {"outputs": Tensor.zeros(OUT).contiguous().realize(), "next_state_feat_q": Tensor.zeros(STATE).contiguous().realize()}}
+    frames = Tensor(self.rng.integers(0, 256, (2, FRAME_SIZE), dtype=np.uint8)).realize()
+    tfm = Tensor(np.tile(np.eye(3, dtype=np.float32), (2, 1, 1))).realize()
+    path = os.path.join(self.tmp.name, "native_tinygrad.pkl")
+    part = compile_jit(native, model_inputs(warp(input_frame=frames, M_inv=tfm).realize()), "output_buffers")
+    extras = {"input_specs": {"new_img": ((2, 6, MODEL_H // 2, MODEL_W // 2), "|u1", "CPU"), "state_feat_q": (STATE, "<f4", "CPU")}
+              | {name: (shape, "<f4", "CPU") for name, shape in HOST.items()},
+              "output_specs": {"outputs": (OUT, "<f4", "CPU"), "next_state_feat_q": (STATE, "<f4", "CPU")},
+              "metadata": {"output_shapes": {"outputs": OUT, "next_state_feat_q": STATE}}}
+    write_bundle(path, {"run": part}, extras)
+
+    fused = load_model(path, cache=False, extend=functools.partial(frame_part, warp_path))
+    assert fused["run"].input_names == ["host", "state_feat_q"]
+    offsets, packed = pack_layout(HOST, FRAME_SIZE)
+    host = np.zeros(packed, np.uint8)
+    inputs = model_inputs(None)
+    for name, shape in ({"tfm": (2, 3, 3)} | HOST).items():
+      np.ndarray(shape, np.float32, host, offsets[name])[...] = tfm.numpy() if name == "tfm" else inputs[name].numpy()
+    host[offsets["frames"]:] = frames.numpy().ravel()
+    fused["run"](host=Tensor(host, device="NPY"), state_feat_q=inputs["state_feat_q"], output_buffers=inputs["output_buffers"])
+
+    expected = model_inputs(warp(input_frame=frames, M_inv=tfm).realize()) | {k: inputs[k] for k in ("state_feat_q",)}
+    native(**expected)
+    for name in ("outputs", "next_state_feat_q"):
+      assert np.array_equal(inputs["output_buffers"][name].numpy(), expected["output_buffers"][name].numpy()), name
+
+  def test_tinygrad_api_the_loader_reads(self):
+    import tinygrad.engine.jit as jit
+    from tinygrad.device import Buffer, TinyELF
+    from tinygrad.helpers import Target
+    from tinygrad.uop.ops import ParamArg, ProgramInfo, UOp
+    fields = {cls: {f.name for f in dataclasses.fields(cls)} for cls in (ProgramInfo, ParamArg, Target)}
+    assert fields[ProgramInfo] >= {"global_size", "local_size", "globals", "outs", "ins", "target"}, "ProgramInfo changed: update helpers._program_uop"
+    assert fields[ParamArg] >= {"slot", "dtype", "name", "image"}, "ParamArg changed: update helpers._program_uop"
+    assert fields[Target] >= {"device", "renderer", "arch"}, "Target changed: update the 'target' key in plain_artifact_compiler.serialize_jit"
+    assert "signature" in getattr(TinyELF, "__annotations__", {}) or hasattr(UOp, "kernel_params"), \
+      "kernel signatures moved: update plain_artifact_compiler.serialize_jit"
+    assert callable(getattr(UOp, "param", None)) and callable(getattr(UOp, "from_buffer", None)), "UOp.param/from_buffer: update helpers.PlainJit"
+    assert callable(getattr(UOp, "store_call", None)) and callable(getattr(UOp, "call", None)), "UOp.call/store_call: update helpers.PlainJit"
+    assert callable(getattr(Buffer, "view", None)) and callable(getattr(Buffer, "ensure_allocated", None)), "Buffer.view: update helpers._upload"
+    assert callable(getattr(jit, "compile_linear", None)) and callable(getattr(jit, "_prepare_jit_inputs", None)), "engine.jit: update helpers._Jit"
+    assert "expected_names" in inspect.signature(jit.CapturedJit).parameters, "CapturedJit: update helpers.PlainJit and plain_artifact_compiler.serialize_jit"

openpilot/sunnypilot/models/manager.py

@@ -7,7 +7,10 @@ See the LICENSE.md file in the root directory for more details.
 
 import asyncio
 import os
+import re
+import threading
 import time
+from concurrent.futures import ThreadPoolExecutor
 
 import requests
 from openpilot.common.params import Params
@@ -22,12 +25,99 @@ from openpilot.sunnypilot.models.helpers import (ACTIVE_BUNDLE_KEYS, get_active_
 
 # (connect, read) seconds. read is per-request inactivity, not a total cap
 DOWNLOAD_TIMEOUT = (30, 30)
+# HuggingFace throttles each connection to ~1-2 MB/s but not the connection count; 12 saturates the device link
+MAX_CONCURRENT_PIECES = 12
+PIECE_SIZE = 8 * 1024 * 1024
+PIECE_ATTEMPTS = 3
+PIECE_BACKOFF = 1.0  # seconds, growing linearly: waits 1 s then 2 s between a piece's three attempts
+TRANSIENT_HTTP_STATUS = frozenset({408, 429, 500, 502, 503, 504})
+REPORT_INTERVAL = 0.5  # seconds between progress publications
+# byte ranges count encoded bytes but iter_content yields decoded ones, so ask for an unencoded body
+NO_COMPRESSION = {"Accept-Encoding": "identity"}
 
 
 class DownloadCancelled(Exception):
   pass
 
 
+def _is_transient(e: BaseException) -> bool:
+  if isinstance(e, requests.HTTPError):
+    return e.response is not None and e.response.status_code in TRANSIENT_HTTP_STATUS
+  return isinstance(e, (requests.ConnectionError, requests.Timeout, requests.exceptions.ChunkedEncodingError))
+
+
+class _Progress:
+  def __init__(self):
+    self._bytes = 0
+    self._lock = threading.Lock()
+
+  def add_bytes(self, n: int) -> None:
+    with self._lock:
+      self._bytes += n
+
+  def snapshot(self) -> int:
+    with self._lock:
+      return self._bytes
+
+
+def _piece_ranges(total: int, piece_size: int) -> list[tuple[int, int]]:
+  return [(start, min(start + piece_size, total)) for start in range(0, total, piece_size)]
+
+
+def _prepare_target(path: str, total: int) -> None:
+  with open(path, "wb") as f:
+    f.truncate(total)
+
+
+def _content_range_total(response: requests.Response) -> int | None:
+  if response.status_code != 206:
+    return None
+  match = re.fullmatch(r"bytes \d+-\d+/(\d+)", response.headers.get("Content-Range", ""))
+  return int(match.group(1)) if match else None
+
+
+def _probe_size(url: str) -> tuple[int, bool]:
+  with requests.get(url, headers={**NO_COMPRESSION, "Range": "bytes=0-0"}, stream=True, timeout=DOWNLOAD_TIMEOUT) as response:
+    response.raise_for_status()
+    total = _content_range_total(response)
+    if total is not None:
+      return total, True
+    return int(response.headers.get("Content-Length", 0)), False
+
+
+def _fetch_piece(url: str, path: str, index: int, start: int, end: int | None, progress: _Progress,
+                 cancel: threading.Event, block_size: int) -> None:
+  headers = {**NO_COMPRESSION, "Range": f"bytes={start}-{end - 1}"} if end is not None else NO_COMPRESSION
+  label = f"{os.path.basename(path)} piece {index}"
+  for attempt in range(PIECE_ATTEMPTS):
+    written = 0
+    try:
+      with requests.get(url, headers=headers, stream=True, timeout=DOWNLOAD_TIMEOUT) as response:
+        response.raise_for_status()
+        if end is not None and response.status_code != 206:
+          raise ValueError(f"server ignored byte range for {label}")
+        with open(path, "r+b") as f:
+          f.seek(start)
+          for block in response.iter_content(chunk_size=block_size):
+            if cancel.is_set():
+              raise DownloadCancelled("Download cancelled")
+            f.write(block)
+            written += len(block)
+            progress.add_bytes(len(block))
+          if end is None:
+            f.truncate()  # unknown length: the body defines the file size
+          if end is not None and written != end - start:
+            raise requests.exceptions.ConnectionError(f"short read: {written} of {end - start} bytes")
+      return
+    except requests.RequestException as e:
+      progress.add_bytes(-written)  # the piece restarts from scratch
+      if not _is_transient(e) or attempt == PIECE_ATTEMPTS - 1:
+        raise
+      cloudlog.warning(f"retrying {label} after {type(e).__name__}: {e}")
+      if cancel.wait(PIECE_BACKOFF * (1 + attempt)):
+        raise DownloadCancelled("Download cancelled") from None
+
+
 class ModelManagerSP:
   """Manages model downloads and status reporting"""
 
@@ -41,7 +131,7 @@ class ModelManagerSP:
     self.source_models: dict[str, list[custom.ModelManagerSP.ModelBundle]] = {}
     self.selected_bundle: custom.ModelManagerSP.ModelBundle = None
     self.active_bundle: custom.ModelManagerSP.ModelBundle = get_active_bundle(self.params, chestnut=self.chestnut_present)
-    self._chunk_size = 128 * 1000  # 128 KB chunks
+    self._block_size = 128 * 1000  # 128 KB network read blocks
     self._download_start_times: dict[str, float] = {}  # Track start time per model
     self._download_ref: bytes | str | None = None
 
@@ -56,7 +146,6 @@ class ModelManagerSP:
     self._download_ref = None
 
   def _sync_artifact_progress(self, source_artifact) -> None:
-    """Mirror download progress to all artifacts sharing the same filename in the selected bundle."""
     if not self.selected_bundle:
       return
     for model in self.selected_bundle.models:
@@ -67,7 +156,6 @@ class ModelManagerSP:
         artifact.downloadProgress.eta = source_artifact.downloadProgress.eta
 
   def _calculate_eta(self, filename: str, progress: float) -> int:
-    """Calculate ETA based on elapsed time and current progress"""
     if filename not in self._download_start_times or progress <= 0:
       return 60  # Default ETA for new downloads
 
@@ -75,39 +163,58 @@ class ModelManagerSP:
     if elapsed_time <= 0:
       return 60
 
-    # If we're at X% after Y seconds, we can estimate total time as (Y / X) * 100
     total_estimated_time = (elapsed_time / progress) * 100
     eta = total_estimated_time - elapsed_time
-
     return max(1, int(eta))  # Return at least 1 second if download is ongoing
 
-  async def _download_file(self, url: str, path: str, model) -> None:
-    """Downloads a file with progress tracking"""
-    self._download_start_times[model.fileName] = time.monotonic()
+  def _set_progress(self, artifact, status, progress: float, eta: int = 0) -> None:
+    artifact.downloadProgress.status = status
+    artifact.downloadProgress.progress = progress
+    artifact.downloadProgress.eta = eta
+    self._sync_artifact_progress(artifact)
+    self._report_status()
 
-    with requests.get(url, stream=True, timeout=DOWNLOAD_TIMEOUT) as response:  # noqa: ASYNC210
-      response.raise_for_status()
-      total_size = int(response.headers.get("content-length", 0))
-      bytes_downloaded = 0
+  def _publish_progress(self, artifact, done_bytes: int, total: int) -> None:
+    # held at 99 until the file passes its hash check
+    progress = min(99.0, done_bytes / total * 100) if total > 0 else 0.0
+    eta = self._calculate_eta(artifact.fileName, progress)
+    self._set_progress(artifact, custom.ModelManagerSP.DownloadStatus.downloading, progress, eta)
 
-      with open(path, 'wb') as f:  # noqa: ASYNC230
-        for chunk in response.iter_content(chunk_size=self._chunk_size):  # type: bytes
-          f.write(chunk)
-          bytes_downloaded += len(chunk)
+  async def _report_until_done(self, tasks: list[asyncio.Future], artifact, progress: _Progress, total: int) -> None:
+    pending = set(tasks)
+    while pending:
+      done, pending = await asyncio.wait(pending, timeout=REPORT_INTERVAL)
+      for task in done:
+        task.result()  # re-raises a worker failure
+      if self._download_interrupted():
+        raise DownloadCancelled("Download cancelled")
+      self._publish_progress(artifact, progress.snapshot(), total)
 
-          if self._download_interrupted():
-            raise DownloadCancelled("Download cancelled")
+  async def _download_file(self, url: str, path: str, artifact) -> None:
+    self._download_start_times[artifact.fileName] = time.monotonic()
+    loop = asyncio.get_running_loop()
 
-          if total_size > 0:
-            progress = (bytes_downloaded / total_size) * 100
-            model.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloading
-            model.downloadProgress.progress = progress
-            model.downloadProgress.eta = self._calculate_eta(model.fileName, progress)
-            self._sync_artifact_progress(model)
-            self._report_status()
+    total, ranged = await loop.run_in_executor(None, _probe_size, url)
+    pieces: list[tuple[int, int | None]] = list(_piece_ranges(total, PIECE_SIZE)) if ranged else [(0, None)]
+    await loop.run_in_executor(None, _prepare_target, path, total)
+    progress = _Progress()
+    cancel = threading.Event()
 
-    # Clean up start time after download completes
-    del self._download_start_times[model.fileName]
+    with ThreadPoolExecutor(max_workers=MAX_CONCURRENT_PIECES) as pool:
+      tasks = [loop.run_in_executor(pool, _fetch_piece, url, path, i, start, end, progress, cancel, self._block_size)
+               for i, (start, end) in enumerate(pieces)]
+      try:
+        await self._report_until_done(tasks, artifact, progress, total)
+      except BaseException as e:
+        cancel.set()
+        if isinstance(e, DownloadCancelled):
+          artifact.downloadProgress.canceling = True
+          self._set_progress(artifact, custom.ModelManagerSP.DownloadStatus.downloading, artifact.downloadProgress.progress)
+        for task in tasks:
+          task.cancel()  # drops queued pieces; running ones see the event at their next block
+        await asyncio.gather(*tasks, return_exceptions=True)
+        raise
+    del self._download_start_times[artifact.fileName]
 
   async def _process_artifact(self, artifact, destination_path: str) -> None:
     if not artifact.downloadUri.uri:
@@ -119,33 +226,30 @@ class ModelManagerSP:
     expected_hash = artifact.downloadUri.sha256
     filename = artifact.fileName
     full_path = os.path.join(destination_path, filename)
+    status = custom.ModelManagerSP.DownloadStatus
+    artifact.downloadProgress.canceling = False
 
     try:
+      self._set_progress(artifact, status.verifying, 0)
       if await verify_file(full_path, expected_hash):
-        artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.cached
-        artifact.downloadProgress.progress = 100
-        artifact.downloadProgress.eta = 0
-        self._sync_artifact_progress(artifact)
-        self._report_status()
+        self._set_progress(artifact, status.cached, 100)
         return
 
+      self._set_progress(artifact, status.downloading, 0)
       await self._download_file(url, full_path, artifact)
+      self._set_progress(artifact, status.verifying, 99)
       if not await verify_file(full_path, expected_hash):
         raise ValueError(f"Hash validation failed for {filename}")
 
-      artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloaded
-      artifact.downloadProgress.progress = 100
-      artifact.downloadProgress.eta = 0
-      self._sync_artifact_progress(artifact)
-      self._report_status()
+      self._set_progress(artifact, status.downloaded, 100)
 
     except DownloadCancelled:
       self._download_start_times.pop(artifact.fileName, None)
-      artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
+      artifact.downloadProgress.status = status.failed
       artifact.downloadProgress.eta = 0
       self._sync_artifact_progress(artifact)
       if self.selected_bundle:
-        self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.failed
+        self.selected_bundle.status = status.failed
       self._report_status()
       raise
 
@@ -153,11 +257,11 @@ class ModelManagerSP:
       cloudlog.error(f"Error downloading {filename}: {str(e)}")
       if os.path.isfile(full_path):  # noqa: ASYNC240
         os.remove(full_path)
-      artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
+      artifact.downloadProgress.status = status.failed
       artifact.downloadProgress.eta = 0
       self._sync_artifact_progress(artifact)
       if self.selected_bundle:
-        self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.failed
+        self.selected_bundle.status = status.failed
       self._report_status()
       self._download_start_times.pop(artifact.fileName, None)
       raise
@@ -272,12 +376,7 @@ class ModelManagerSP:
         rk.keep_time()
 
   def clear_model_cache(self) -> None:
-    """
-    Clears the model cache directory of all files except those in the active model bundle.
-    """
-
-    # Get list of files used by both slots' selected bundles (either may become
-    # the truly active bundle depending on hardware availability)
+    """ Clears the model cache directory of all files except those in the active model bundle."""
     active_files = []
     for source in ACTIVE_BUNDLE_KEYS:
       if selected_bundle := get_selected_bundle(self.params, source):
@@ -285,7 +384,6 @@ class ModelManagerSP:
           if model.artifact.fileName:
             active_files.append(model.artifact.fileName)
 
-    # Remove all files except active ones
     model_dir = Paths.model_root()
     try:
       for filename in os.listdir(model_dir):

openpilot/system/ui/sunnypilot/widgets/download_status.py

@@ -57,7 +57,7 @@ class DownloadStatusAction(ItemAction):
     self.icon: str | None = None
     self.icon_color: rl.Color | None = None
     self._font = gui_app.font(FontWeight.NORMAL)
-    # raw progress arrives in steps, one per 128KB chunk the manager publishes
+    # raw progress arrives in steps, one per progress report from the manager
     self._progress = FirstOrderFilter(0.0, 0.5, 1 / gui_app.target_fps)
     # integrated per frame; (t * speed) % span jumps whenever the fill width changes
     self._sweep = 0.0