chore: sync tinygrad (#1680)

* chore: sync tinygrad

Runs great in sim. now we need to rebuild some models

* oops forgot to unblock this after testing

* helpers

* oh yeah

* latest tg

* this wont do anything empriically

* reduce complexity

* okay lint

* Update tinygrad_runner.py

* Update modeld.py

* Update build-all-tinygrad-models.yaml

* tinygrad bump

* Update modeld.py

* Update tinygrad_runner.py

* bump

* Update SConscript

* Update SConscript

* com

* Update fetcher.py

* Update helpers.py

* Update model_runner.py

* Update model_runner.py

---------

Co-authored-by: Jason Wen <haibin.wen3@gmail.com>

.github/workflows/build-all-tinygrad-models.yaml

@@ -140,7 +140,7 @@ jobs:
         run: |
           echo '${{ needs.setup.outputs.model_matrix }}' > matrix.json
           built=(); while IFS= read -r line; do built+=("$line"); done < <(
-            ls output | sed -E 's/^model-//' | sed -E 's/-[0-9]+$//' | sed -E 's/ \([^)]*\)//' | awk '{gsub(/^ +| +$/, ""); print}'
+            find output -maxdepth 1 -name 'model-*' -printf "%f\n" | sed -E 's/^model-//' | sed -E 's/-[0-9]+$//' | sed -E 's/ \([^)]*\)//' | awk '{gsub(/^ +| +$/, ""); print}'
           )
           jq -c --argjson built "$(printf '%s\n' "${built[@]}" | jq -R . | jq -s .)" \
             'map(select(.display_name as $n | ($built | index($n | gsub("^ +| +$"; "")) | not)))' matrix.json > retry_matrix.json
@@ -168,6 +168,7 @@ jobs:
     if: ${{ !cancelled() && (needs.get_and_build.result != 'failure' || needs.retry_get_and_build.result == 'success' || (needs.retry_failed_models.outputs.retry_matrix != '[]' && needs.retry_failed_models.outputs.retry_matrix != '')) }}
     runs-on: ubuntu-latest
     strategy:
+      fail-fast: false
       max-parallel: 1
       matrix:
         model: ${{ fromJson(needs.setup.outputs.model_matrix) }}

.gitmodules

@@ -15,7 +15,7 @@
   url = https://github.com/commaai/teleoprtc
 [submodule "tinygrad"]
   path = tinygrad_repo
-  url = https://github.com/commaai/tinygrad.git
+  url = https://github.com/sunnypilot/tinygrad.git
 [submodule "sunnypilot/neural_network_data"]
   path = sunnypilot/neural_network_data
   url = https://github.com/sunnypilot/neural-network-data.git

selfdrive/modeld/SConscript

@@ -50,7 +50,7 @@ def tg_compile(flags, model_name):
 # Compile small models
 for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
   flags = {
-    'larch64': 'DEV=QCOM',
+    'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 IMAGE=2 JIT_BATCH_SIZE=0',
     'Darwin': f'DEV=CPU HOME={os.path.expanduser("~")} IMAGE=0', # tinygrad calls brew which needs a $HOME in the env
   }.get(arch, 'DEV=CPU CPU_LLVM=1 IMAGE=0')
   tg_compile(flags, model_name)

sunnypilot/modeld_v2/SConscript

@@ -60,7 +60,7 @@ def tg_compile(flags, model_name):
 for model_name in ['supercombo', 'driving_vision', 'driving_off_policy', 'driving_policy']:
   if File(f"models/{model_name}.onnx").exists():
     flags = {
-      'larch64': 'DEV=QCOM',
+      'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 IMAGE=2 JIT_BATCH_SIZE=0',
       'Darwin': f'DEV=CPU HOME={os.path.expanduser("~")} IMAGE=0', # tinygrad calls brew which needs a $HOME in the env
     }.get(arch, 'DEV=CPU CPU_LLVM=1 IMAGE=0')
     tg_compile(flags, model_name)

sunnypilot/modeld_v2/modeld.py

@@ -1,4 +1,11 @@
 #!/usr/bin/env python3
+import os
+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'
+  os.environ['AMD_IFACE'] = 'USB'
 import time
 import numpy as np
 import cereal.messaging as messaging

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-docs/refs/heads/gh-pages/docs/driving_models_v11.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-docs/refs/heads/gh-pages/docs/driving_models_v14.json"
 
   def __init__(self, params: Params):
     self.params = params

sunnypilot/models/helpers.py

@@ -19,8 +19,8 @@ from openpilot.system.hardware.hw import Paths
 from pathlib import Path
 
 # see the README.md for more details on the model selector versioning
-CURRENT_SELECTOR_VERSION = 13
-REQUIRED_MIN_SELECTOR_VERSION = 12
+CURRENT_SELECTOR_VERSION = 14
+REQUIRED_MIN_SELECTOR_VERSION = 14
 
 USE_ONNX = os.getenv('USE_ONNX', PC)
 

sunnypilot/models/runners/model_runner.py

@@ -1,9 +1,7 @@
-import os
 from abc import abstractmethod, ABC
 
 import numpy as np
 from openpilot.sunnypilot.models.helpers import get_active_bundle
-from openpilot.system.hardware import TICI
 from openpilot.sunnypilot.models.runners.constants import NumpyDict, ShapeDict, CLMemDict, FrameDict, Model, SliceDict, SEND_RAW_PRED
 from openpilot.system.hardware.hw import Paths
 import pickle
@@ -11,15 +9,6 @@ import pickle
 CUSTOM_MODEL_PATH = Paths.model_root()
 
 
-# Set QCOM environment variable for TICI devices, potentially enabling hardware acceleration
-USBGPU = "USBGPU" in os.environ
-if USBGPU:
-  os.environ['DEV'] = 'AMD'
-  os.environ['AMD_IFACE'] = 'USB'
-else:
-  os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
-
-
 class ModelData:
   """
   Stores metadata and configuration for a specific machine learning model.

sunnypilot/models/runners/tinygrad/tinygrad_runner.py

@@ -51,7 +51,7 @@ class TinygradRunner(ModelRunner, SupercomboTinygrad, PolicyTinygrad, VisionTiny
     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_st_vars_dtype_device[idx]
+      info = self.model_run.captured.expected_input_info[idx]
       self.input_to_dtype[name] = info[2]  # dtype
       self.input_to_device[name] = info[3]  # device
 
@@ -84,7 +84,7 @@ class TinygradRunner(ModelRunner, SupercomboTinygrad, PolicyTinygrad, VisionTiny
 
   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()
+    outputs = self.model_run(**self.inputs).numpy().flatten()
     return self._parse_outputs(outputs)
 
   def _parse_outputs(self, model_outputs: np.ndarray) -> NumpyDict: