ref

* Revert "modeld: fix capnp memory leak (#38117)"

This reverts commit 0f17a98793.

* Revert "fix memory leak from pycapnp 2.2+ (#38001)"

This reverts commit 294cb687f6.

* downgrade

* agnos 18.4

.gitattributes

@@ -11,13 +11,4 @@
 *.wav filter=lfs diff=lfs merge=lfs -text
 
 selfdrive/car/tests/test_models_segs.txt filter=lfs diff=lfs merge=lfs -text
-system/hardware/tici/updater_weston filter=lfs diff=lfs merge=lfs -text
-system/hardware/tici/updater_magic filter=lfs diff=lfs merge=lfs -text
-third_party/**/*.a filter=lfs diff=lfs merge=lfs -text
-third_party/**/*.so filter=lfs diff=lfs merge=lfs -text
-third_party/**/*.so.* filter=lfs diff=lfs merge=lfs -text
-third_party/**/*.dylib filter=lfs diff=lfs merge=lfs -text
-third_party/acados/*/t_renderer filter=lfs diff=lfs merge=lfs -text
-third_party/qt5/larch64/bin/lrelease filter=lfs diff=lfs merge=lfs -text
-third_party/qt5/larch64/bin/lupdate filter=lfs diff=lfs merge=lfs -text
-third_party/catch2/include/catch2/catch.hpp filter=lfs diff=lfs merge=lfs -text
+system/hardware/tici/updater filter=lfs diff=lfs merge=lfs -text

.github/ISSUE_TEMPLATE/enhancement.md

@@ -1,8 +0,0 @@
----
-name: Enhancement
-about: For openpilot enhancement suggestions
-title: ''
-labels: 'enhancement'
-assignees: ''
----
-

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

@@ -120,6 +120,7 @@ jobs:
     with:
       upstream_branch: ${{ matrix.model.ref }}
       custom_name: ${{ matrix.model.display_name }}
+      is_20hz: ${{ matrix.model.is_20hz }}
       recompiled_dir: ${{ needs.setup.outputs.recompiled_dir }}
       json_version: ${{ needs.setup.outputs.json_version }}
     secrets: inherit
@@ -157,6 +158,7 @@ jobs:
     with:
       upstream_branch: ${{ matrix.model.ref }}
       custom_name: ${{ matrix.model.display_name }}
+      is_20hz: ${{ matrix.model.is_20hz }}
       recompiled_dir: ${{ needs.setup.outputs.recompiled_dir }}
       json_version: ${{ needs.setup.outputs.json_version }}
       artifact_suffix: -retry

.github/workflows/build-single-tinygrad-model.yaml

@@ -24,6 +24,11 @@ on:
         required: false
         type: string
         default: ''
+      is_20hz:
+        description: 'Is this a 20Hz model'
+        required: false
+        type: boolean
+        default: true
       bypass_push:
         description: 'Bypass pushing to GitLab for build-all'
         required: false
@@ -39,6 +44,11 @@ on:
         description: 'Custom name for the model (no date, only name)'
         required: false
         type: string
+      is_20hz:
+        description: 'Is this a 20Hz model'
+        required: false
+        type: boolean
+        default: true
       recompiled_dir:
         description: 'Existing recompiled directory number (e.g. 3 for recompiled3)'
         required: true
@@ -82,7 +92,7 @@ jobs:
     with:
       upstream_branch: ${{ inputs.upstream_branch }}
       custom_name: ${{ inputs.custom_name || inputs.upstream_branch }}
-      is_20hz: true
+      is_20hz: ${{ inputs.is_20hz }}
       artifact_suffix: ${{ inputs.artifact_suffix }}
     secrets: inherit
 

.github/workflows/cereal_validation.yaml

@@ -23,56 +23,43 @@ env:
   CI: 1
 
 jobs:
-  generate_cereal_artifact:
-    name: Generate cereal validation artifacts
-    runs-on: ubuntu-24.04
-    steps:
-      - uses: actions/checkout@v6
-        with:
-          submodules: true
-      - run: ./tools/op.sh setup
-      - name: Build openpilot
-        run: scons -j$(nproc) cereal
-      - name: Dump sunnypilot schema
-        run: |
-          export PYTHONPATH=${{ github.workspace }}
-          python3 cereal/messaging/tests/validate_sp_cereal_upstream.py -g -f schema.json
-      - name: 'Prepare artifact'
-        run: |
-          mkdir -p "cereal/messaging/tests/cereal_validations"
-          cp cereal/messaging/tests/validate_sp_cereal_upstream.py "cereal/messaging/tests/cereal_validations/validate_sp_cereal_upstream.py"
-          cp schema.json "cereal/messaging/tests/cereal_validations/schema.json"
-      - name: 'Upload Artifact'
-        uses: actions/upload-artifact@v4
-        with:
-          name: cereal_validations
-          path: cereal/messaging/tests/cereal_validations
-
   validate_cereal_with_upstream:
     name: Validate cereal with Upstream
     runs-on: ubuntu-24.04
-    needs: generate_cereal_artifact
     steps:
-      - name: Checkout sunnypilot
+      - name: Checkout sunnypilot cereal
         uses: actions/checkout@v6
-      - name: Checkout upstream openpilot
+        with:
+          sparse-checkout: cereal
+
+      - name: Init sunnypilot opendbc submodule
+        run: git submodule update --init --depth 1 opendbc_repo
+
+      - name: Checkout upstream openpilot cereal
         uses: actions/checkout@v6
         with:
           repository: 'commaai/openpilot'
-          path: openpilot
-          submodules: true
+          path: upstream_openpilot
+          sparse-checkout: cereal
           ref: "refs/heads/master"
-      - run: ./tools/op.sh setup
-      - name: Build openpilot
-        working-directory: openpilot
-        run: scons -j$(nproc) cereal
-      - name: Download build artifacts
-        uses: actions/download-artifact@v4
-        with:
-          name: cereal_validations
-          path: openpilot/cereal/messaging/tests/cereal_validations
-      - name: 'Validate sunnypilot schema against upstream'
+
+      - name: Init upstream opendbc submodule
+        working-directory: upstream_openpilot
+        run: git submodule update --init --depth 1 opendbc_repo
+
+      - name: Install uv
+        run: pip install uv
+
+      - name: Generate sunnypilot schema
         run: |
-          export PYTHONPATH=${{ github.workspace }}/openpilot
-          chmod +x openpilot/cereal/messaging/tests/cereal_validations/validate_sp_cereal_upstream.py
-          python3 openpilot/cereal/messaging/tests/cereal_validations/validate_sp_cereal_upstream.py -r -f openpilot/cereal/messaging/tests/cereal_validations/schema.json
+          PYCAPNP_VER=$(python3 -c "import re; m=re.search(r'name = \"pycapnp\"\nversion = \"([^\"]+)\"', open('uv.lock').read()); print(m.group(1))")
+          uv run --isolated --with "pycapnp==${PYCAPNP_VER}" \
+            python3 cereal/messaging/tests/validate_sp_cereal_upstream.py \
+            -g -f /tmp/sp_schema.json --cereal-dir cereal
+
+      - name: Validate against upstream
+        run: |
+          PYCAPNP_VER=$(python3 -c "import re; m=re.search(r'name = \"pycapnp\"\nversion = \"([^\"]+)\"', open('uv.lock').read()); print(m.group(1))")
+          uv run --isolated --with "pycapnp==${PYCAPNP_VER}" \
+            python3 cereal/messaging/tests/validate_sp_cereal_upstream.py \
+            -r -f /tmp/sp_schema.json --cereal-dir upstream_openpilot/cereal

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

@@ -164,18 +164,54 @@ jobs:
           source /etc/profile
           export UV_PROJECT_ENVIRONMENT=${HOME}/venv
           export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
-          export PYTHONPATH="${PYTHONPATH}:${{ env.TINYGRAD_PATH }}"
+          export PYTHONPATH="${PYTHONPATH}:${{ env.TINYGRAD_PATH }}:${{ github.workspace }}"
 
-          # Loop through all .onnx files
+          COMPILE_MODELD="${{ github.workspace }}/sunnypilot/modeld_v2/compile_modeld.py"
+          MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
+          CAMERA_RES=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
+          TG_FLAGS="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+
+          # Generate metadata for all ONNX files
           find "${{ env.MODELS_DIR }}" -maxdepth 1 -name '*.onnx' | while IFS= read -r onnx_file; do
-            base_name=$(basename "$onnx_file" .onnx)
-            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 python3 "${{ env.MODELS_DIR }}/../get_model_metadata.py" "$onnx_file" || true
+            echo "Generating metadata: $onnx_file"
+            env ${TG_FLAGS} python3 "${{ env.MODELS_DIR }}/../get_model_metadata.py" "$onnx_file" || true
           done
 
+          # Detect model type and build compile args
+          VISION_ONNX="${{ env.MODELS_DIR }}/driving_vision.onnx"
+          POLICY_ONNX="${{ env.MODELS_DIR }}/driving_policy.onnx"
+          OFF_POLICY_ONNX="${{ env.MODELS_DIR }}/driving_off_policy.onnx"
+          ON_POLICY_ONNX="${{ env.MODELS_DIR }}/driving_on_policy.onnx"
+          SUPERCOMBO_ONNX="${{ env.MODELS_DIR }}/supercombo.onnx"
+
+          MODEL_TYPE="" ONNX_ARGS="" OUTPUT_NAME=""
+          if [ -f "$VISION_ONNX" ]; then
+            ONNX_ARGS="--vision-onnx $VISION_ONNX"
+            if [ -f "$ON_POLICY_ONNX" ] && [ -f "$OFF_POLICY_ONNX" ]; then
+              MODEL_TYPE=vision_multi_policy
+              ONNX_ARGS="$ONNX_ARGS --off-policy-onnx $OFF_POLICY_ONNX --on-policy-onnx $ON_POLICY_ONNX"
+            elif [ -f "$OFF_POLICY_ONNX" ] && [ -f "$POLICY_ONNX" ]; then
+              MODEL_TYPE=vision_multi_policy
+              ONNX_ARGS="$ONNX_ARGS --policy-onnx $POLICY_ONNX --off-policy-onnx $OFF_POLICY_ONNX"
+            elif [ -f "$POLICY_ONNX" ]; then
+              MODEL_TYPE=vision_policy
+              ONNX_ARGS="$ONNX_ARGS --policy-onnx $POLICY_ONNX"
+            fi
+          elif [ -f "$SUPERCOMBO_ONNX" ]; then
+            MODEL_TYPE=supercombo
+            ONNX_ARGS="--supercombo-onnx $SUPERCOMBO_ONNX"
+          fi
+
+          if [ -n "$MODEL_TYPE" ]; then
+            echo "Detected: $MODEL_TYPE -> driving_tinygrad.pkl"
+            env ${TG_FLAGS} python3 "$COMPILE_MODELD" \
+              --model-type $MODEL_TYPE \
+              --model-size $MODEL_SIZE \
+              --camera-resolutions $CAMERA_RES \
+              $ONNX_ARGS \
+              --output "${{ env.MODELS_DIR }}/driving_tinygrad.pkl"
+          fi
+
       - name: Validate Model Outputs
         run: |
           source /etc/profile
@@ -194,6 +230,8 @@ jobs:
           rsync -avm \
             --include='*.dlc' \
             --include='*.pkl' \
+            --include='*.chunk*' \
+            --include='*.chunkmanifest' \
             --include='*.onnx' \
             --exclude='*' \
             --delete-excluded \

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

@@ -215,8 +215,8 @@ jobs:
             --exclude='**/SConstruct' \
             --exclude='**/SConscript' \
             --exclude='**/.venv/' \
-            --exclude='selfdrive/modeld/models/driving_vision.onnx' \
-            --exclude='selfdrive/modeld/models/driving_policy.onnx' \
+            --exclude='selfdrive/modeld/models/*.onnx*' \
+            --exclude='sunnypilot/modeld*/models/*.onnx*' \
             --exclude='third_party/*x86*' \
             --exclude='third_party/*Darwin*' \
             --delete-excluded \

.github/workflows/tests.yaml

@@ -123,7 +123,7 @@ jobs:
         submodules: true
     - run: ./tools/op.sh setup
     - name: Build openpilot
-      run: scons -j$(nproc)
+      run: scons
     - name: Run unit tests
       timeout-minutes: ${{ contains(runner.name, 'nsc') && 2 || 999 }}
       run: |
@@ -147,7 +147,7 @@ jobs:
         submodules: true
     - run: ./tools/op.sh setup
     - name: Build openpilot
-      run: scons -j$(nproc)
+      run: scons
     - name: Run replay
       timeout-minutes: ${{ contains(runner.name, 'nsc') && 2 || 20 }}
       continue-on-error: ${{ github.ref == 'refs/heads/master' }}
@@ -179,7 +179,7 @@ jobs:
         repository: commaai/ci-artifacts
         ssh-key: ${{ secrets.CI_ARTIFACTS_DEPLOY_KEY }}
         path: ${{ github.workspace }}/ci-artifacts
-    - name: Push refs
+    - name: Prepare refs
       if: github.repository == 'commaai/openpilot' && github.ref == 'refs/heads/master'
       working-directory: ${{ github.workspace }}/ci-artifacts
       run: |
@@ -191,7 +191,13 @@ jobs:
         echo "${{ github.sha }}" > ref_commit
         git add .
         git commit -m "process-replay refs for ${{ github.repository }}@${{ github.sha }}" || echo "No changes to commit"
-        git push origin process-replay --force
+    - name: Push refs
+      if: github.repository == 'commaai/openpilot' && github.ref == 'refs/heads/master'
+      uses: nick-fields/retry@7152eba30c6575329ac0576536151aca5a72780e
+      with:
+        timeout_minutes: 2
+        max_attempts: 3
+        command: cd ${{ github.workspace }}/ci-artifacts && git push origin process-replay --force
     - name: Run regen
       if: false
       timeout-minutes: 4
@@ -214,7 +220,7 @@ jobs:
         submodules: true
     - run: ./tools/op.sh setup
     - name: Build openpilot
-      run: scons -j$(nproc)
+      run: scons
     - name: Driving test
       timeout-minutes: 2
       run: |
@@ -235,7 +241,7 @@ jobs:
           submodules: true
       - run: ./tools/op.sh setup
       - name: Build openpilot
-        run: scons -j$(nproc)
+        run: scons
       - name: Create UI Report
         run: |
           source selfdrive/test/setup_xvfb.sh

.gitignore

@@ -44,6 +44,7 @@ bin/
 config.json
 compile_commands.json
 compare_runtime*.html
+selfdrive/modeld/models/tg_input_devices.json
 
 # build artifacts
 docs_site/

.gitmodules

@@ -4,6 +4,7 @@
 [submodule "opendbc"]
   path = opendbc_repo
   url = https://github.com/sunnypilot/opendbc.git
+  branch = tn
 [submodule "msgq"]
   path = msgq_repo
   url = https://github.com/commaai/msgq.git

.vscode/settings.json

@@ -21,7 +21,6 @@
     "common/**",
     "selfdrive/**",
     "system/**",
-    "third_party/**",
     "tools/**",
   ]
 }

CHANGELOG.md

@@ -1,4 +1,7 @@
-sunnypilot Version 2026.001.000 (2026-03-xx)
+sunnypilot Version 2026.002.000 (2026-xx-xx)
+========================
+
+sunnypilot Version 2026.001.000 (2026-05-06)
 ========================
 * What's Changed (sunnypilot/sunnypilot)
   * Complete rewrite of the user interface from Qt C++ to Raylib Python
@@ -66,6 +69,64 @@ sunnypilot Version 2026.001.000 (2026-03-xx)
   * Pause Lateral Control with Blinker: Post-Blinker Delay by @CHaucke89
   * SCC-V: Use p97 for predicted lateral accel by @yasu-oh
   * Controls: Support for Torque Lateral Control v0 Tune by @sunnyhaibin
+  * [TIZI/TICI] ui: ensure null checks for `CarParams` and `CarParamsSP` by @sunnyhaibin
+  * [TIZI/TICI] ui: use `vCruiseCluster` and `vEgoCluster` for SLA `preActive` by @sunnyhaibin
+  * Fix display of values when using use_float_scaling by @CHaucke89
+  * models: fix default & index "0" by @nayan8teen
+  * [TIZI/TICI] visuals: Improved speed limit by @angaz
+  * ICBM: ensure button timers update on disable to clear stale presses by @jamesmikesell
+  * [TIZI/TICI] ui: simplify Smart Cruise Control text rendering by @sunnyhaibin
+  * controlsd: fix steer_limited_by_safety not updating under MADS by @zephleggett
+  * soundd: trigger timeout warning during MADS lateral-only by @zephleggett
+  * pandad: flasher for Rivian long upgrade module by @lukasloetkolben
+  * modeld_v2: tinygrad transformation warp by @Discountchubbs
+  * tools: block `manage_sunnylinkd` in sim startup script by @sunnyhaibin
+  * [MICI] ui: need superclass `_render` in `HudRendererSP` by @sunnyhaibin
+  * [TIZI/TICI] ui: Speed Limit Assist active status by @sunnyhaibin
+  * ui: reimplement "Screen Off" option to Onroad Brightness by @sunnyhaibin
+  * ui: don't hide steering wheel when blindspot disabled by @royjr
+  * ui: Speed Limit Assist `preActive` improvements  by @sunnyhaibin
+  * ui: consolidate Speed Limit Assist `preActive` status rendering by @sunnyhaibin
+  * [MICI] ui: Speed Limit Assist `preActive` status by @sunnyhaibin
+  * sunnypilot modeld: remove thneed modeld by @Discountchubbs
+  * modeld_v2: decouple planplus scaling from accel by @Discountchubbs
+  * sunnylink: Handle exceptions in `getParamsAllKeysV1` to log crashes by @devtekve
+  * [TIZI/TICI] ui: Developer UI cleanup by @sunnyhaibin
+  * [TIZI/TICI] ui: dynamic alert size by @nayan8teen
+  * i18n(fr): Add French translations by @didlawowo
+  * Toyota: Stop and Go Hack (Alpha) by @sunnyhaibin
+  * ui: `AlertFadeAnimator` for longitudinal-related statuses by @sunnyhaibin
+  * pandad: gate unsupported pandas before flashing by @sunnyhaibin
+  * Rivian: Flash xnor's Longitudinal Upgrade Kit prior supported panda check by @lukasloetkolben
+  * [TIZI/TICI] ui: add back gate steering arc behind toggle by @sunnyhaibin
+  * ui: gate Onroad Brightness Delay on readiness by @sunnyhaibin
+  * ui: add new timer options for Onroad Brightness Delay by @sunnyhaibin
+  * [TIZI/TICI] ui: branch switcher is always available by @sunnyhaibin
+  * pandad: always prioritize internal panda by @sunnyhaibin
+  * sunnylinkd: fetch compressed params schema by @sunnyhaibin
+  * sunnypilot locationd: remove unused car_ekf filter by @sunnyhaibin
+  * modeld_v2: update deprecated temporalPose ref by @sunnyhaibin
+  * NNLC: restore pre-v1 PID gains in torque extension by @mmmorks
+  * MADS safety: enable heartbeat and lateral controls mismatch checks by @sunnyhaibin
+  * [MICI] ui: models panel enhancements by @nayan8teen
+  * [TIZI/TICI] ui: fix unintended selection while scrolling in TreeOptionDialog by @TheSecurityDev
+  * tools: script for video concatenation by @Discountchubbs
+  * tools: profile memory usage by @Discountchubbs
+  * [TIZI/TICI] ui: remove per-frame param sync by @sunnyhaibin
+  * [MICI] ui: always offroad by @nayan8teen
+  * controls: always default Torque Lateral Control to v0 Tune by @sunnyhaibin
+  * Revert "controls: always default Torque Lateral Control to v0 Tune" by @sunnyhaibin
+  * Reapply "controls: always default Torque Lateral Control to v0 Tune" (#1806) by @sunnyhaibin
+  * [MICI] ui: add sunnylink info & connectivity check by @nayan8teen
+  * sunnylink: Remove unused API endpoint by @devtekve
+  * DM: wheel touch enforcement in MADS by @sunnyhaibin
+  * torque: show static override values in Dev UI & gate `useParams` on custom torque tune by @sunnyhaibin
+  * MADS: suppress espActive event when long is not engaged by @sunnyhaibin
+  * sunnylink: SDUI by @sunnyhaibin
+  * [MICI] ui: align upstream changes with sunnypilot settings buttons by @nayan8teen
+  * ui: fix cellular toggles by @AmyJeanes
+  * sunnylink: switch athena domain by @devtekve
+  * Platform List: dynamically migrate CarPlatformBundle by @sunnyhaibin
 * What's Changed (sunnypilot/opendbc)
   * Honda: DBC for Accord 9th Generation by @mvl-boston
   * FCA: update tire stiffness values for `RAM_HD` by @dparring
@@ -84,12 +145,25 @@ sunnypilot Version 2026.001.000 (2026-03-xx)
   * Honda: add missing `GasInterceptor` messages to Taiwan Odyssey DBC by @mvl-boston
   * GM: remove `CHEVROLET_EQUINOX_NON_ACC_3RD_GEN` from `dashcamOnly` by @sunnyhaibin
   * GM: remove `CHEVROLET_BOLT_NON_ACC_2ND_GEN` from `dashcamOnly` by @sunnyhaibin
+  * Hyundai Longitudinal: deprecate ramp update for dynamic tune by @Discountchubbs
+  * Rivian: long upgrade messages on bus 1 by @lukasloetkolben
+  * Toyota: Stop and Go Hack (Alpha) by @sunnyhaibin
+  * Toyota: gate Smart DSU behind Alpha Longitudinal by @sunnyhaibin
+  * Toyota: Gas Interceptor always set `standstill_req` by @sunnyhaibin
+  * MADS safety: dedicated `controls_allowed_lateral` by @sunnyhaibin
+  * Platform List: include community supported platforms by @sunnyhaibin
 * New Contributors (sunnypilot/sunnypilot)
   * @TheSecurityDev made their first contribution in "ui: fix sidebar scroll in UI screenshots"
   * @zikeji made their first contribution in "sunnylink: block remote modification of SSH key parameters"
   * @Candy0707 made their first contribution in "[TIZI/TICI] ui: Fix misaligned turn signals and blindspot indicators with sidebar"
   * @CHaucke89 made their first contribution in "Pause Lateral Control with Blinker: Post-Blinker Delay"
   * @yasu-oh made their first contribution in "SCC-V: Use p97 for predicted lateral accel"
+  * @angaz made their first contribution in "[TIZI/TICI] visuals: Improved speed limit"
+  * @jamesmikesell made their first contribution in "ICBM: ensure button timers update on disable to clear stale presses"
+  * @zephleggett made their first contribution in "controlsd: fix steer_limited_by_safety not updating under MADS"
+  * @lukasloetkolben made their first contribution in "pandad: flasher for Rivian long upgrade module"
+  * @didlawowo made their first contribution in "i18n(fr): Add French translations"
+  * @mmmorks made their first contribution in "NNLC: restore pre-v1 PID gains in torque extension"
 * New Contributors (sunnypilot/opendbc)
   * @AmyJeanes made their first contribution in "Tesla: Fix stock LKAS being blocked when MADS is enabled"
   * @mvl-boston made their first contribution in "Honda: Update Clarity brake to renamed DBC message name"
@@ -99,6 +173,20 @@ sunnypilot Version 2026.001.000 (2026-03-xx)
   * @royjr made their first contribution in "HKG: add KIA_FORTE_2019_NON_SCC fingerprint"
   * @ssysm made their first contribution in "Tesla: remove `TESLA_MODEL_X` from `dashcamOnly`"
 * Full Changelog: https://github.com/sunnypilot/sunnypilot/compare/v2025.002.000...v2026.001.000
+************************
+* Synced with commaai's openpilot (v0.11.1)
+  * master commit c001f3c9b490a80e69539f0af6022f6e07ceb721 (April 16, 2026)
+* New driver monitoring model
+* Improved image processing pipeline for driver camera
+* Rivian R1S and R1T 2025 support thanks to lukasloetkolben!
+* New driving model #36798
+  * Fully trained using a learned simulator
+  * Improved longitudinal performance in Experimental mode
+* Reduce comma four standby power usage by 77% to 52 mW
+* Kia K7 2017 support thanks to royjr!
+* Lexus LS 2018 support thanks to Hacheoy!
+* Improved inter-process communication memory efficiency
+* comma four support
 
 sunnypilot Version 2025.002.000 (2025-11-06)
 ========================

Jenkinsfile

@@ -166,8 +166,8 @@ node {
   env.GIT_BRANCH = checkout(scm).GIT_BRANCH
   env.GIT_COMMIT = checkout(scm).GIT_COMMIT
 
-  def excludeBranches = ['__nightly', 'devel', 'devel-staging', 'release3', 'release3-staging',
-                         'release-tici', 'release-tizi', 'release-tizi-staging', 'release-mici-staging', 'testing-closet*', 'hotfix-*']
+  def excludeBranches = ['__nightly', 'devel', 'devel-staging',
+                         'release-tizi', 'release-tizi-staging', 'release-mici', 'release-mici-staging', 'testing-closet*', 'hotfix-*']
   def excludeRegex = excludeBranches.join('|').replaceAll('\\*', '.*')
 
   if (env.BRANCH_NAME != 'master' && !env.BRANCH_NAME.contains('__jenkins_loop_')) {
@@ -179,7 +179,7 @@ node {
   try {
     if (env.BRANCH_NAME == 'devel-staging') {
       deviceStage("build release-tizi-staging", "tizi-needs-can", [], [
-        step("build release-tizi-staging", "RELEASE_BRANCH=release-tizi-staging $SOURCE_DIR/release/build_release.sh && git push -f origin release-tizi-staging:release-mici-staging"),
+        step("build release-tizi-staging", "RELEASE_BRANCH=release-tizi-staging,release-mici-staging $SOURCE_DIR/release/build_release.sh"),
       ])
     }
 
@@ -247,7 +247,6 @@ node {
           step("test pandad loopback", "pytest selfdrive/pandad/tests/test_pandad_loopback.py"),
           step("test pandad spi", "pytest selfdrive/pandad/tests/test_pandad_spi.py"),
           step("test amp", "pytest system/hardware/tici/tests/test_amplifier.py"),
-          step("test qcomgpsd", "pytest system/qcomgpsd/tests/test_qcomgpsd.py", [diffPaths: ["system/qcomgpsd/"]]),
         ])
       },
 

RELEASES.md

@@ -1,7 +1,13 @@
-Version 0.11.1 (2026-04-22)
+Version 0.11.2 (2026-06-15)
+========================
+* Acceleration Profile: select Eco, Normal, or Sport personality to shape acceleration briskness and gentle (ACC) deceleration
+
+Version 0.11.1 (2026-05-18)
 ========================
 * New driver monitoring model
 * Improved image processing pipeline for driver camera
+* Improved thermal policy for comma four
+* Acura MDX 2022-24 support thanks to mvl-boston!
 * Rivian R1S and R1T 2025 support thanks to lukasloetkolben!
 
 Version 0.11.0 (2026-03-17)

SConstruct

@@ -10,25 +10,28 @@ import numpy as np
 import SCons.Errors
 from SCons.Defaults import _stripixes
 
+TICI = os.path.isfile('/TICI')
+
 SCons.Warnings.warningAsException(True)
 
 Decider('MD5-timestamp')
 
-SetOption('num_jobs', max(1, int(os.cpu_count()/2)))
+SetOption('num_jobs', max(1, int(os.cpu_count()/(1 if "CI" in os.environ else 2))))
 
 AddOption('--ccflags', action='store', type='string', default='', help='pass arbitrary flags over the command line')
 AddOption('--verbose', action='store_true', default=False, help='show full build commands')
+release = not os.path.exists(File('#.gitattributes').abspath) # file absent on release branch, see release_files.py
 AddOption('--minimal',
           action='store_false',
           dest='extras',
-          default=os.path.exists(File('#.gitattributes').abspath), # minimal by default on release branch (where there's no LFS)
+          default=(not TICI and not release),
           help='the minimum build to run openpilot. no tests, tools, etc.')
 
 # Detect platform
 arch = subprocess.check_output(["uname", "-m"], encoding='utf8').rstrip()
 if platform.system() == "Darwin":
   arch = "Darwin"
-elif arch == "aarch64" and os.path.isfile('/TICI'):
+elif arch == "aarch64" and TICI:
   arch = "larch64"
 assert arch in [
   "larch64",  # linux tici arm64
@@ -37,8 +40,14 @@ assert arch in [
   "Darwin",   # macOS arm64 (x86 not supported)
 ]
 
-pkg_names = ['bzip2', 'capnproto', 'eigen', 'ffmpeg', 'libjpeg', 'libyuv', 'ncurses', 'zeromq', 'zstd']
+pkg_names = ['acados', 'bzip2', 'capnproto', 'catch2', 'eigen', 'ffmpeg', 'json11', 'libjpeg', 'libyuv', 'ncurses', 'zeromq', 'zstd']
 pkgs = [importlib.import_module(name) for name in pkg_names]
+acados = pkgs[pkg_names.index('acados')]
+acados_include_dirs = [
+  acados.INCLUDE_DIR,
+  os.path.join(acados.INCLUDE_DIR, "blasfeo", "include"),
+  os.path.join(acados.INCLUDE_DIR, "hpipm", "include"),
+]
 
 
 # ***** enforce a whitelist of system libraries *****
@@ -82,10 +91,10 @@ def _libflags(target, source, env, for_signature):
 env = Environment(
   ENV={
     "PATH": os.environ['PATH'],
-    "PYTHONPATH": Dir("#").abspath + ':' + Dir(f"#third_party/acados").abspath,
-    "ACADOS_SOURCE_DIR": Dir("#third_party/acados").abspath,
-    "ACADOS_PYTHON_INTERFACE_PATH": Dir("#third_party/acados/acados_template").abspath,
-    "TERA_PATH": Dir("#").abspath + f"/third_party/acados/{arch}/t_renderer"
+    "PYTHONPATH": Dir("#").abspath,
+    "ACADOS_SOURCE_DIR": acados.DIR,
+    "ACADOS_PYTHON_INTERFACE_PATH": acados.TEMPLATE_DIR,
+    "TERA_PATH": acados.TERA_PATH
   },
   CCFLAGS=[
     "-g",
@@ -105,22 +114,14 @@ env = Environment(
   CPPPATH=[
     "#",
     "#msgq",
-    "#third_party",
-    "#third_party/json11",
-    "#third_party/linux/include",
-    "#third_party/acados/include",
-    "#third_party/acados/include/blasfeo/include",
-    "#third_party/acados/include/hpipm/include",
-    "#third_party/catch2/include",
+    acados_include_dirs,
     [x.INCLUDE_DIR for x in pkgs],
   ],
   LIBPATH=[
     "#common",
     "#msgq_repo",
-    "#third_party",
     "#selfdrive/pandad",
     "#rednose/helpers",
-    f"#third_party/acados/{arch}/lib",
     [x.LIB_DIR for x in pkgs],
   ],
   RPATH=[],
@@ -174,16 +175,6 @@ if not GetOption('verbose'):
   ):
     env[f"{action}COMSTR"] = f"  [{short}] $TARGET"
 
-# progress output
-node_interval = 5
-node_count = 0
-def progress_function(node):
-  global node_count
-  node_count += node_interval
-  sys.stderr.write("progress: %d\n" % node_count)
-if os.environ.get('SCONS_PROGRESS'):
-  Progress(progress_function, interval=node_interval)
-
 # ********** Cython build environment **********
 envCython = env.Clone()
 envCython["CPPPATH"] += [sysconfig.get_paths()['include'], np.get_include()]
@@ -199,14 +190,24 @@ else:
 np_version = SCons.Script.Value(np.__version__)
 Export('envCython', 'np_version')
 
-Export('env', 'arch')
+Export('env', 'arch', 'acados', 'release')
 
 # Setup cache dir
 default_cache_dir = '/data/scons_cache' if arch == "larch64" else '/tmp/scons_cache'
 cache_dir = ARGUMENTS.get('cache_dir', default_cache_dir)
+cache_size_limit = 4e9 if "CI" in os.environ else 2e9
 CacheDir(cache_dir)
 Clean(["."], cache_dir)
 
+def prune_cache_dir(target=None, source=None, env=None):
+  cache_files = sorted((os.path.join(root, f) for root, _, files in os.walk(cache_dir) for f in files), key=os.path.getmtime)
+  cache_size = sum(os.path.getsize(f) for f in cache_files)
+  for f in cache_files:
+    if cache_size < cache_size_limit:
+      break
+    cache_size -= os.path.getsize(f)
+    os.unlink(f)
+
 # ********** start building stuff **********
 
 # Build common module
@@ -242,9 +243,6 @@ SConscript([
 if arch == "larch64":
   SConscript(['system/camerad/SConscript'])
 
-# Build openpilot
-SConscript(['third_party/SConscript'])
-
 # Build selfdrive
 SConscript([
   'selfdrive/pandad/SConscript',
@@ -257,8 +255,8 @@ SConscript([
 
 SConscript(['sunnypilot/SConscript'])
 
-# Build tools
-if arch != "larch64":
+# Build desktop-only tools
+if GetOption('extras') and arch != "larch64":
   SConscript([
     'tools/replay/SConscript',
     'tools/cabana/SConscript',
@@ -267,3 +265,37 @@ if arch != "larch64":
 
 
 env.CompilationDatabase('compile_commands.json')
+
+# progress output
+def count_scons_nodes(nodes):
+  seen = set()
+  stack = list(nodes)
+
+  while stack:
+    node = stack.pop().disambiguate()
+    if node in seen:
+      continue
+    seen.add(node)
+    executor = node.get_executor()
+    if executor is not None:
+      stack += executor.get_all_prerequisites() + executor.get_all_children()
+
+  return len(seen)
+
+progress_interval = 5
+progress_count = 0
+progress_total = max(1, count_scons_nodes(env.arg2nodes(BUILD_TARGETS or [Dir('.')], env.fs.Entry)))
+
+def progress_function(node):
+  global progress_count
+  if progress_count >= progress_total:
+    return
+  progress_count = min(progress_count + progress_interval, progress_total)
+  progress = round(100. * progress_count / progress_total, 1)
+  sys.stderr.write("\rBuilding: %5.1f%%" % progress if sys.stderr.isatty() else "progress: %.1f\n" % progress)
+  if progress == 100. and sys.stderr.isatty():
+    sys.stderr.write("\n")
+  sys.stderr.flush()
+
+Progress(progress_function, interval=progress_interval)
+AddPostAction(BUILD_TARGETS or [Dir('.')], prune_cache_dir)

cereal/custom.capnp

@@ -194,6 +194,7 @@ struct LongitudinalPlanSP @0xf35cc4560bbf6ec2 {
   aTarget @5 :Float32;
   events @6 :List(OnroadEventSP.Event);
   e2eAlerts @7 :E2eAlerts;
+  acceleration @8 :Acceleration;
 
   struct DynamicExperimentalControl {
     state @0 :DynamicExperimentalControlState;
@@ -296,6 +297,20 @@ struct LongitudinalPlanSP @0xf35cc4560bbf6ec2 {
     greenLightAlert @0 :Bool;
     leadDepartAlert @1 :Bool;
   }
+
+  struct Acceleration {
+    personality @0 :AccelerationPersonality;
+    enabled @1 :Bool;
+    maxAccel @2 :Float32;
+    tFollowOffset @3 :Float32;
+    jerkFactorScale @4 :Float32;
+  }
+
+  enum AccelerationPersonality {
+    eco @0;
+    normal @1;
+    sport @2;
+  }
 }
 
 struct OnroadEventSP @0xda96579883444c35 {
@@ -342,6 +357,7 @@ struct OnroadEventSP @0xda96579883444c35 {
     speedLimitChanged @21;
     speedLimitPending @22;
     e2eChime @23;
+    laneChangeRoadEdge @24;
   }
 }
 
@@ -448,6 +464,8 @@ struct LiveMapDataSP @0xf416ec09499d9d19 {
 
 struct ModelDataV2SP @0xa1680744031fdb2d {
   laneTurnDirection @0 :TurnDirection;
+  leftLaneChangeEdgeBlock @1 :Bool;
+  rightLaneChangeEdgeBlock @2 :Bool;
 
   enum TurnDirection {
     none @0;

cereal/log.capnp

@@ -273,11 +273,7 @@ struct GPSNMEAData {
   nmea @2 :Text;
 }
 
-# android sensor_event_t
 struct SensorEventData {
-  version @0 :Int32;
-  sensor @1 :Int32;
-  type @2 :Int32;
   timestamp @3 :Int64;
 
   union {
@@ -296,7 +292,10 @@ struct SensorEventData {
 
   struct SensorVec {
     v @0 :List(Float32);
-    status @1 :Int8;
+
+    deprecated :group {
+      status @1 :Int8;
+    }
   }
 
   enum SensorSource {
@@ -314,7 +313,11 @@ struct SensorEventData {
     mmc5603nj @11;
   }
 
+  # formerly based on android sensor_event_t
   deprecated :group {
+    version @0 :Int32;
+    sensor @1 :Int32;
+    type @2 :Int32;
     uncalibrated @10 :Bool;
   }
 }
@@ -457,10 +460,10 @@ struct DeviceState @0xa4d8b5af2aa492eb {
   }
 
   enum ThermalStatus {
-    green @0;
-    yellow @1;
-    red @2;
-    danger @3;
+    ok @0;
+    warmDEPRECATED @1;
+    overheated @2;
+    critical @3;
   }
 
   enum NetworkType {
@@ -2054,16 +2057,15 @@ struct DriverStateV2 {
     facePosition @2 :List(Float32);
     facePositionStd @3 :List(Float32);
     faceProb @4 :Float32;
-    eyesVisibleProb @14 :Float32;
-    eyesClosedProb @15 :Float32;
+    leftEyeProb @5 :Float32;
+    rightEyeProb @6 :Float32;
+    leftBlinkProb @7 :Float32;
+    rightBlinkProb @8 :Float32;
+    sunglassesProb @9 :Float32;
     phoneProb @13 :Float32;
+    sleepProb @14 :Float32;
 
     deprecated :group {
-      leftEyeProb @5 :Float32;
-      rightEyeProb @6 :Float32;
-      leftBlinkProb @7 :Float32;
-      rightBlinkProb @8 :Float32;
-      sunglassesProb @9 :Float32;
       notReadyProb @12 :List(Float32);
       occludedProb @10 :Float32;
       readyProb @11 :List(Float32);
@@ -2076,7 +2078,7 @@ struct DriverStateV2 {
   }
 }
 
-struct DriverMonitoringState @0xb83cda094a1da284 {
+struct DriverMonitoringStateDEPRECATED @0xb83cda094a1da284 {
   events @18 :List(OnroadEvent);
   faceDetected @1 :Bool;
   isDistracted @2 :Bool;
@@ -2104,6 +2106,75 @@ struct DriverMonitoringState @0xb83cda094a1da284 {
   }
 }
 
+struct DriverMonitoringState {
+  lockout @0 :Bool;
+  alertCountLockoutPercent @1 :Int8;
+  alertTimeLockoutPercent @2 :Int8;
+
+  alwaysOn @3 :Bool;
+  alwaysOnLockout @4 :Bool;
+
+  alertLevel @5 :AlertLevel;
+  activePolicy @6 :MonitoringPolicy;
+  isRHD @7 :Bool;
+  rhdCalibration @8 :CalibrationState;
+
+  visionPolicyState @9 :VisionPolicyState;
+  wheeltouchPolicyState @10 :WheeltouchPolicyState;
+
+  enum AlertLevel {
+    # ordinal must match the name to prevent bugs
+    # comparing against the raw ordinal value
+    none @0;
+    one @1;
+    two @2;
+    three @3;
+  }
+
+  enum MonitoringPolicy {
+    wheeltouch @0;
+    vision @1;
+  }
+
+  struct VisionPolicyState {
+    awarenessPercent @0 :Int8;
+    awarenessStep @1 :Float32;
+    isDistracted @2 :Bool;
+    distractedTypes @3 :DistractedTypes;
+
+    faceDetected @4 :Bool;
+    pose @5 :Pose;
+    wheeltouchFallbackPercent @6 :Int8;
+    uncertainOffroadAlertPercent @7 :Int8;
+
+    struct DistractedTypes {
+      pose @0: Bool;
+      eye @1: Bool;
+      phone @2: Bool;
+    }
+
+    struct Pose {
+      pitch @0 :Float32;
+      yaw @1 :Float32;
+      pitchCalib @2 :CalibrationState;
+      yawCalib @3 :CalibrationState;
+      calibrated @4 :Bool;
+      uncertainty @5 :Float32;
+    }
+  }
+
+  struct WheeltouchPolicyState {
+    awarenessPercent @0 :Int8;
+    awarenessStep @1 :Float32;
+    driverInteracting @2 :Bool;
+  }
+
+  struct CalibrationState {
+    calibratedPercent @0 :Int8;
+    offset @1 :Float32;
+  }
+}
+
 struct Boot {
   wallTimeNanos @0 :UInt64;
   pstore @4 :Map(Text, Data);
@@ -2229,7 +2300,8 @@ struct Sentinel {
 }
 
 struct UIDebug {
-  drawTimeMillis @0 :Float32;
+  cpuTimeMillis @0 :Float32;
+  frameTimeMillis @1 :Float32;
 }
 
 struct ManagerState {
@@ -2377,7 +2449,6 @@ struct Event {
     boot @60 :Boot;
 
     # ********** openpilot daemon msgs **********
-    gpsNMEA @3 :GPSNMEAData;
     can @5 :List(CanData);
     controlsState @7 :ControlsState;
     selfdriveState @130 :SelfdriveState;
@@ -2402,7 +2473,6 @@ struct Event {
     qcomGnss @31 :QcomGnss;
     gpsLocationExternal @48 :GpsLocationData;
     gpsLocation @21 :GpsLocationData;
-    gnssMeasurements @91 :GnssMeasurements;
     liveParameters @61 :LiveParametersData;
     liveTorqueParameters @94 :LiveTorqueParametersData;
     liveDelay @146 : LiveDelayData;
@@ -2410,7 +2480,7 @@ struct Event {
     thumbnail @66: Thumbnail;
     onroadEvents @134: List(OnroadEvent);
     carParams @69: Car.CarParams;
-    driverMonitoringState @71: DriverMonitoringState;
+    driverMonitoringState @151 :DriverMonitoringState;
     livePose @129 :LivePose;
     modelV2 @75 :ModelDataV2;
     drivingModelData @128 :DrivingModelData;
@@ -2436,7 +2506,6 @@ struct Event {
     # systems stuff
     androidLog @20 :AndroidLogEntry;
     managerState @78 :ManagerState;
-    uploaderState @79 :UploaderState;
     procLog @33 :ProcLog;
     clocks @35 :Clocks;
     deviceState @6 :DeviceState;
@@ -2446,12 +2515,6 @@ struct Event {
     # touch frame
     touch @135 :List(Touch);
 
-    # navigation
-    navInstruction @82 :NavInstruction;
-    navRoute @83 :NavRoute;
-    navThumbnail @84: Thumbnail;
-    mapRenderState @105: MapRenderState;
-
     # UI services
     uiDebug @102 :UIDebug;
 
@@ -2553,5 +2616,13 @@ struct Event {
     gyroscope2DEPRECATED @100 :SensorEventData;
     accelerometer2DEPRECATED @101 :SensorEventData;
     temperatureSensor2DEPRECATED @123 :SensorEventData;
+    driverMonitoringStateDEPRECATED @71 :DriverMonitoringStateDEPRECATED;
+    gpsNMEADEPRECATED @3 :GPSNMEAData;
+    uploaderStateDEPRECATED @79 :UploaderState;
+    navInstructionDEPRECATED @82 :NavInstruction;
+    navRouteDEPRECATED @83 :NavRoute;
+    navThumbnailDEPRECATED @84 :Thumbnail;
+    gnssMeasurementsDEPRECATED @91 :GnssMeasurements;
+    mapRenderStateDEPRECATED @105: MapRenderState;
   }
 }

cereal/messaging/__init__.py

@@ -259,11 +259,11 @@ class PubMaster:
     self.sock[s].send(dat)
 
   def wait_for_readers_to_update(self, s: str, timeout: int, dt: float = 0.05) -> bool:
-    for _ in range(int(timeout*(1./dt))):
-      if self.sock[s].all_readers_updated():
-        return True
-      time.sleep(dt)
-    return False
+    try:
+      self.sock[s].wait_for_readers(timeout=timeout, interval=dt)
+      return True
+    except TimeoutError:
+      return False
 
   def all_readers_updated(self, s: str) -> bool:
-    return self.sock[s].all_readers_updated()  # type: ignore
+    return self.sock[s].all_readers_updated()

cereal/messaging/tests/test_messaging.py

@@ -30,7 +30,7 @@ def zmq_sleep(t=1):
 
 # TODO: this should take any capnp struct and returrn a msg with random populated data
 def random_carstate():
-  fields = ["vEgo", "aEgo", "brake", "steeringAngleDeg"]
+  fields = ["vEgo", "aEgo", "steeringTorque", "steeringAngleDeg"]
   msg = messaging.new_message("carState")
   cs = msg.carState
   for f in fields:

cereal/messaging/tests/validate_sp_cereal_upstream.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 
 import argparse
 import json
+import os
 import sys
 from typing import Any
 
@@ -104,8 +105,15 @@ def collect_schema(root: Any) -> dict[str, dict]:
   return structs
 
 
-def dump_schema(path: str) -> None:
-  from cereal import log
+def load_log(cereal_dir: str) -> Any:
+  import capnp
+  cereal_dir = os.path.abspath(cereal_dir)
+  capnp.remove_import_hook()
+  return capnp.load(os.path.join(cereal_dir, "log.capnp"), imports=[cereal_dir])
+
+
+def dump_schema(cereal_dir: str, path: str) -> None:
+  log = load_log(cereal_dir)
   payload = {
     "root": hex_id(log.Event.schema.node.id),
     "structs": collect_schema(log.Event.schema),
@@ -206,8 +214,8 @@ def load_peer(path: str) -> dict:
     return json.load(handle)
 
 
-def run_read(peer_path: str) -> int:
-  from cereal import log
+def run_read(cereal_dir: str, peer_path: str) -> int:
+  log = load_log(cereal_dir)
   peer_dump = load_peer(peer_path)
   local_dump = {
     "root": hex_id(log.Event.schema.node.id),
@@ -235,16 +243,13 @@ def main() -> int:
   mode.add_argument("-g", "--generate", action="store_true", help="dump local schema to JSON")
   mode.add_argument("-r", "--read", action="store_true", help="load peer JSON and diff against local")
   parser.add_argument("-f", "--file", default="schema.json", help="JSON file path (default: schema.json)")
+  parser.add_argument("--cereal-dir", required=True, help="path to cereal directory containing log.capnp")
   args = parser.parse_args()
 
-  try:
-    if args.generate:
-      dump_schema(args.file)
-      return 0
-    return run_read(args.file)
-  except ImportError as exc:
-    print(f"error: cannot import cereal ({exc}). did scons build cereal?")
-    return 2
+  if args.generate:
+    dump_schema(args.cereal_dir, args.file)
+    return 0
+  return run_read(args.cereal_dir, args.file)
 
 
 if __name__ == "__main__":

cereal/services.py

@@ -24,10 +24,7 @@ _services: dict[str, tuple] = {
   # note: the "EncodeIdx" packets will still be in the log
   "gyroscope": (True, 104., 104),
   "accelerometer": (True, 104., 104),
-  "magnetometer": (True, 25.),
-  "lightSensor": (True, 100., 100),
   "temperatureSensor": (True, 2., 200),
-  "gpsNMEA": (True, 9.),
   "deviceState": (True, 2., 1),
   "touch": (True, 20., 1),
   "can": (True, 100., 2053, QueueSize.BIG),  # decimation gives ~3 msgs in a full segment
@@ -56,7 +53,6 @@ _services: dict[str, tuple] = {
   "gpsLocation": (True, 1., 1),
   "ubloxGnss": (True, 10.),
   "qcomGnss": (True, 2.),
-  "gnssMeasurements": (True, 10., 10),
   "clocks": (True, 0.1, 1),
   "ubloxRaw": (True, 20.),
   "livePose": (True, 20., 4),
@@ -75,10 +71,6 @@ _services: dict[str, tuple] = {
   "drivingModelData": (True, 20., 10),
   "modelV2": (True, 20., None, QueueSize.BIG),
   "managerState": (True, 2., 1),
-  "uploaderState": (True, 0., 1),
-  "navInstruction": (True, 1., 10),
-  "navRoute": (True, 0.),
-  "navThumbnail": (True, 0.),
   "qRoadEncodeIdx": (False, 20.),
   "userBookmark": (True, 0., 1),
   "soundPressure": (True, 10., 10),
@@ -114,8 +106,6 @@ _services: dict[str, tuple] = {
   "livestreamRoadEncodeData": (False, 20., None, QueueSize.MEDIUM),
   "livestreamDriverEncodeData": (False, 20., None, QueueSize.MEDIUM),
   "customReservedRawData0": (True, 0.),
-  "customReservedRawData1": (True, 0.),
-  "customReservedRawData2": (True, 0.),
 }
 SERVICE_LIST = {name: Service(*vals) for
                 idx, (name, vals) in enumerate(_services.items())}

common/file_chunker.py

@@ -1,3 +1,5 @@
+#!/usr/bin/env python3
+import sys
 import math
 import os
 from pathlib import Path
@@ -10,10 +12,13 @@ def get_chunk_name(name, idx, num_chunks):
 def get_manifest_path(name):
   return f"{name}.chunkmanifest"
 
-def get_chunk_paths(path, file_size):
-  num_chunks = math.ceil(file_size / CHUNK_SIZE)
+def _chunk_paths(path, num_chunks):
   return [get_manifest_path(path)] + [get_chunk_name(path, i, num_chunks) for i in range(num_chunks)]
 
+def get_chunk_targets(path, file_size):
+  num_chunks = math.ceil(file_size / CHUNK_SIZE)
+  return _chunk_paths(path, num_chunks)
+
 def chunk_file(path, targets):
   manifest_path, *chunk_paths = targets
   with open(path, 'rb') as f:
@@ -26,6 +31,13 @@ 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)
 
 def read_file_chunked(path):
   manifest_path = get_manifest_path(path)
@@ -35,3 +47,9 @@ def read_file_chunked(path):
   if os.path.isfile(path):
     return Path(path).read_bytes()
   raise FileNotFoundError(path)
+
+
+if __name__ == "__main__":
+  path = sys.argv[1]
+  chunk_paths = get_chunk_targets(path, os.path.getsize(path))
+  chunk_file(path, chunk_paths)

common/model.h

@@ -1 +0,0 @@
-#define DEFAULT_MODEL "POP model (Default)"

common/parameterized.py

@@ -1,8 +1,13 @@
+import re
 import sys
 import pytest
 import inspect
 
 
+def _to_safe_name(s):
+  return re.sub(r"[^a-zA-Z0-9_]+", "_", str(s)).strip("_")
+
+
 class parameterized:
   @staticmethod
   def expand(cases):
@@ -34,7 +39,9 @@ def parameterized_class(attrs, input_list=None):
   def decorator(cls):
     globs = sys._getframe(1).f_globals
     for i, params in enumerate(params_list):
-      name = f"{cls.__name__}_{i}"
+      # append sanitized string param values so pytest -k can filter by them
+      suffix = "_".join(filter(None, (_to_safe_name(v) for v in params.values() if isinstance(v, str))))
+      name = f"{cls.__name__}_{i}" + (f"_{suffix}" if suffix else "")
       new_cls = type(name, (cls,), dict(params))
       new_cls.__module__ = cls.__module__
       new_cls.__test__ = True  # override inherited False so pytest collects this subclass

common/params.py

@@ -14,6 +14,6 @@ if __name__ == "__main__":
   if len(sys.argv) == 3:
     val = sys.argv[2]
     print(f"SET: {key} = {val}")
-    params.put(key, val)
+    params.put(key, val, block=True)
   elif len(sys.argv) == 2:
     print(f"GET: {key} = {params.get(key)}")

common/params_keys.h

@@ -80,6 +80,7 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"LiveDelay", {PERSISTENT | BACKUP, BYTES}},
     {"LiveParameters", {PERSISTENT, JSON}},
     {"LiveParametersV2", {PERSISTENT, BYTES}},
+    {"LivestreamEncoderBitrate", {CLEAR_ON_MANAGER_START | DONT_LOG, INT}},
     {"LiveTorqueParameters", {PERSISTENT | DONT_LOG, BYTES}},
     {"LocationFilterInitialState", {PERSISTENT, BYTES}},
     {"LateralManeuverMode", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
@@ -103,8 +104,6 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"OnroadCycleRequested", {CLEAR_ON_MANAGER_START, BOOL}},
     {"OpenpilotEnabledToggle", {PERSISTENT | BACKUP, BOOL, "1"}},
     {"PandaHeartbeatLost", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
-    {"PandaSomResetTriggered", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
-    {"PandaSignatures", {CLEAR_ON_MANAGER_START, BYTES}},
     {"PrimeType", {PERSISTENT, INT}},
     {"RecordAudio", {PERSISTENT | BACKUP, BOOL}},
     {"RecordAudioFeedback", {PERSISTENT | BACKUP, BOOL, "0"}},
@@ -132,6 +131,8 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"UpdaterLastFetchTime", {PERSISTENT, TIME}},
     {"UptimeOffroad", {PERSISTENT, FLOAT, "0.0"}},
     {"UptimeOnroad", {PERSISTENT, FLOAT, "0.0"}},
+    {"UsbGpuPresent", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
+    {"UsbGpuCompiled", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, BOOL}},
     {"Version", {PERSISTENT, STRING}},
 
     // --- sunnypilot params --- //
@@ -178,12 +179,19 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"QuickBootToggle", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"QuietMode", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"RainbowMode", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"RoadEdgeLaneChangeEnabled", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"RocketFuel", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"ShowAdvancedControls", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"ShowTurnSignals", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"StandstillTimer", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"TrueVEgoUI", {PERSISTENT | BACKUP, BOOL, "0"}},
 
+    // toyota specific params
+    {"ToyotaAutoHold", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"ToyotaEnhancedBsm", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"ToyotaTSS2Long", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"ToyotaDriveMode", {PERSISTENT | BACKUP, BOOL, "0"}},
+
     // MADS params
     {"Mads", {PERSISTENT | BACKUP, BOOL, "1"}},
     {"MadsMainCruiseAllowed", {PERSISTENT | BACKUP, BOOL, "1"}},
@@ -204,6 +212,7 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     // sunnylink params
     {"EnableSunnylinkUploader", {PERSISTENT | BACKUP, BOOL}},
     {"LastSunnylinkPingTime", {CLEAR_ON_MANAGER_START, INT}},
+    {"ParamsVersion", {PERSISTENT, INT}},
     {"SunnylinkCache_Roles", {PERSISTENT, STRING}},
     {"SunnylinkCache_Users", {PERSISTENT, STRING}},
     {"SunnylinkDongleId", {PERSISTENT, STRING}},
@@ -224,6 +233,10 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
     {"ToyotaStopAndGoHack", {PERSISTENT | BACKUP, BOOL, "0"}},
 
     {"DynamicExperimentalControl", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"AccelPersonalityEnabled", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"AccelPersonality", {PERSISTENT | BACKUP, INT, "1"}},
+    {"RadarDistance", {PERSISTENT | BACKUP, BOOL, "0"}},
+    {"SmoothBraking", {PERSISTENT | BACKUP, BOOL, "0"}},
     {"BlindSpot", {PERSISTENT | BACKUP, BOOL, "0"}},
 
     // sunnypilot model params

common/params_pyx.pyx

@@ -142,33 +142,28 @@ cdef class Params:
     cdef ParamKeyType t = self.p.getKeyType(k)
     return ensure_bytes(self.python2cpp(type(dat), t, dat, key))
 
-  def put(self, key, dat):
+  def put(self, key, dat, bool block = False):
     """
-    Warning: This function blocks until the param is written to disk!
+    Warning: block=True blocks until the param is written to disk!
     In very rare cases this can take over a second, and your code will hang.
-    Use the put_nonblocking, put_bool_nonblocking in time sensitive code, but
-    in general try to avoid writing params as much as possible.
+    Use block=False in time sensitive code, but in general try to avoid
+    writing params as much as possible.
     """
     cdef string k = self.check_key(key)
     cdef string dat_bytes = self._put_cast(key, dat)
     with nogil:
-      self.p.put(k, dat_bytes)
+      if block:
+        self.p.put(k, dat_bytes)
+      else:
+        self.p.putNonBlocking(k, dat_bytes)
 
-  def put_bool(self, key, bool val):
+  def put_bool(self, key, bool val, bool block = False):
     cdef string k = self.check_key(key)
     with nogil:
-      self.p.putBool(k, val)
-
-  def put_nonblocking(self, key, dat):
-    cdef string k = self.check_key(key)
-    cdef string dat_bytes = self._put_cast(key, dat)
-    with nogil:
-      self.p.putNonBlocking(k, dat_bytes)
-
-  def put_bool_nonblocking(self, key, bool val):
-    cdef string k = self.check_key(key)
-    with nogil:
-      self.p.putBoolNonBlocking(k, val)
+      if block:
+        self.p.putBool(k, val)
+      else:
+        self.p.putBoolNonBlocking(k, val)
 
   def remove(self, key):
     cdef string k = self.check_key(key)

common/realtime.py

@@ -28,6 +28,11 @@ class Priority:
   CTRL_HIGH = 53
 
 
+def drop_realtime() -> None:
+  if sys.platform == 'linux' and not PC:
+    os.sched_setscheduler(0, os.SCHED_OTHER, os.sched_param(0))
+
+
 def set_core_affinity(cores: list[int]) -> None:
   if sys.platform == 'linux' and not PC:
     os.sched_setaffinity(0, cores)

common/swaglog.cc

@@ -11,7 +11,7 @@
 
 #include <zmq.h>
 #include <stdarg.h>
-#include "third_party/json11/json11.hpp"
+#include "json11/json11.hpp"
 #include "common/version.h"
 #include "system/hardware/hw.h"
 

common/tests/test_params.py

@@ -12,17 +12,17 @@ class TestParams:
     self.params = Params()
 
   def test_params_put_and_get(self):
-    self.params.put("DongleId", "cb38263377b873ee")
+    self.params.put("DongleId", "cb38263377b873ee", block=True)
     assert self.params.get("DongleId") == "cb38263377b873ee"
 
   def test_params_non_ascii(self):
     st = b"\xe1\x90\xff"
-    self.params.put("CarParams", st)
+    self.params.put("CarParams", st, block=True)
     assert self.params.get("CarParams") == st
 
   def test_params_get_cleared_manager_start(self):
-    self.params.put("CarParams", b"test")
-    self.params.put("DongleId", "cb38263377b873ee")
+    self.params.put("CarParams", b"test", block=True)
+    self.params.put("DongleId", "cb38263377b873ee", block=True)
     assert self.params.get("CarParams") == b"test"
 
     undefined_param = self.params.get_param_path(uuid.uuid4().hex)
@@ -36,15 +36,15 @@ class TestParams:
     assert not os.path.isfile(undefined_param)
 
   def test_params_two_things(self):
-    self.params.put("DongleId", "bob")
-    self.params.put("AthenadPid", 123)
+    self.params.put("DongleId", "bob", block=True)
+    self.params.put("AthenadPid", 123, block=True)
     assert self.params.get("DongleId") == "bob"
     assert self.params.get("AthenadPid") == 123
 
   def test_params_get_block(self):
     def _delayed_writer():
       time.sleep(0.1)
-      self.params.put("CarParams", b"test")
+      self.params.put("CarParams", b"test", block=True)
     threading.Thread(target=_delayed_writer).start()
     assert self.params.get("CarParams") is None
     assert self.params.get("CarParams", block=True) == b"test"
@@ -57,10 +57,10 @@ class TestParams:
       self.params.get_bool("swag")
 
     with pytest.raises(UnknownKeyName):
-      self.params.put("swag", "abc")
+      self.params.put("swag", "abc", block=True)
 
     with pytest.raises(UnknownKeyName):
-      self.params.put_bool("swag", True)
+      self.params.put_bool("swag", True, block=True)
 
   def test_remove_not_there(self):
     assert self.params.get("CarParams") is None
@@ -71,23 +71,23 @@ class TestParams:
     self.params.remove("IsMetric")
     assert not self.params.get_bool("IsMetric")
 
-    self.params.put_bool("IsMetric", True)
+    self.params.put_bool("IsMetric", True, block=True)
     assert self.params.get_bool("IsMetric")
 
-    self.params.put_bool("IsMetric", False)
+    self.params.put_bool("IsMetric", False, block=True)
     assert not self.params.get_bool("IsMetric")
 
-    self.params.put("IsMetric", True)
+    self.params.put("IsMetric", True, block=True)
     assert self.params.get_bool("IsMetric")
 
-    self.params.put("IsMetric", False)
+    self.params.put("IsMetric", False, block=True)
     assert not self.params.get_bool("IsMetric")
 
   def test_put_non_blocking_with_get_block(self):
     q = Params()
     def _delayed_writer():
       time.sleep(0.1)
-      Params().put_nonblocking("CarParams", b"test")
+      Params().put("CarParams", b"test")
     threading.Thread(target=_delayed_writer).start()
     assert q.get("CarParams") is None
     assert q.get("CarParams", True) == b"test"
@@ -96,7 +96,7 @@ class TestParams:
     q = Params()
     def _delayed_writer():
       time.sleep(0.1)
-      Params().put_bool_nonblocking("CarParams", True)
+      Params().put_bool("CarParams", True)
     threading.Thread(target=_delayed_writer).start()
     assert q.get("CarParams") is None
     assert q.get("CarParams", True) == b"1"
@@ -123,19 +123,19 @@ class TestParams:
 
   def test_params_get_type(self):
     # json
-    self.params.put("ApiCache_FirehoseStats", {"a": 0})
+    self.params.put("ApiCache_FirehoseStats", {"a": 0}, block=True)
     assert self.params.get("ApiCache_FirehoseStats") == {"a": 0}
 
     # int
-    self.params.put("BootCount", 1441)
+    self.params.put("BootCount", 1441, block=True)
     assert self.params.get("BootCount") == 1441
 
     # bool
-    self.params.put("AdbEnabled", True)
+    self.params.put("AdbEnabled", True, block=True)
     assert self.params.get("AdbEnabled")
     assert isinstance(self.params.get("AdbEnabled"), bool)
 
     # time
     now = datetime.datetime.now(datetime.UTC)
-    self.params.put("InstallDate", now)
+    self.params.put("InstallDate", now, block=True)
     assert self.params.get("InstallDate") == now

common/tests/test_swaglog.cc

@@ -7,7 +7,7 @@
 #include "common/util.h"
 #include "common/version.h"
 #include "system/hardware/hw.h"
-#include "third_party/json11/json11.hpp"
+#include "json11/json11.hpp"
 
 #include "sunnypilot/common/version.h"
 

common/utils.py

@@ -48,7 +48,7 @@ def sudo_write(val: str, path: str) -> None:
 
 def sudo_read(path: str) -> str:
   try:
-    return subprocess.check_output(f"sudo cat {path}", shell=True, encoding='utf8').strip()
+    return subprocess.check_output(["sudo", "cat", "--", path], encoding='utf8').strip()
   except Exception:
     return ""
 

common/version.h

@@ -1 +1 @@
-#define COMMA_VERSION "0.11.1"
+#define COMMA_VERSION "0.11.2"

conftest.py

@@ -7,25 +7,19 @@ from openpilot.common.prefix import OpenpilotPrefix
 from openpilot.system.manager import manager
 from openpilot.system.hardware import TICI, HARDWARE
 
-# TODO: pytest-cpp doesn't support FAIL, and we need to create test translations in sessionstart
-# pending https://github.com/pytest-dev/pytest-cpp/pull/147
+# these are heavy CI-only tests, invoked explicitly in .github/workflows/tests.yaml
 collect_ignore = [
   "selfdrive/test/process_replay/test_processes.py",
   "selfdrive/test/process_replay/test_regen.py",
+  # tinygrad JIT has process-global state. Other test files import modeld → tinygrad,
+  # which corrupts JIT captures for test_warp.py in the same process. Run separately in CI.
+  "sunnypilot/modeld_v2/tests/test_warp.py",
 ]
 collect_ignore_glob = [
   "selfdrive/debug/*.py",
-  "selfdrive/modeld/*.py",
-  "sunnypilot/modeld*/*.py",
 ]
 
 
-def pytest_sessionstart(session):
-  # TODO: fix tests and enable test order randomization
-  if session.config.pluginmanager.hasplugin('randomly'):
-    session.config.option.randomly_reorganize = False
-
-
 @pytest.hookimpl(hookwrapper=True, trylast=True)
 def pytest_runtest_call(item):
   # ensure we run as a hook after capturemanager's
@@ -97,15 +91,3 @@ def pytest_collection_modifyitems(config, items):
       class_property_name = item.get_closest_marker('xdist_group_class_property').args[0]
       class_property_value = getattr(item.cls, class_property_name)
       item.add_marker(pytest.mark.xdist_group(class_property_value))
-
-
-@pytest.hookimpl(trylast=True)
-def pytest_configure(config):
-  config_line = "xdist_group_class_property: group tests by a property of the class that contains them"
-  config.addinivalue_line("markers", config_line)
-
-  config_line = "nocapture: don't capture test output"
-  config.addinivalue_line("markers", config_line)
-
-  config_line = "shared_download_cache: share download cache between tests"
-  config.addinivalue_line("markers", config_line)

docs/SAFETY.md

@@ -4,8 +4,8 @@ openpilot is an Adaptive Cruise Control (ACC) and Automated Lane Centering (ALC)
 Like other ACC and ALC systems, openpilot is a failsafe passive system and it requires the
 driver to be alert and to pay attention at all times.
 
-In order to enforce driver alertness, openpilot includes a driver monitoring feature
-that alerts the driver when distracted.
+To assist the driver in maintaining alertness, openpilot includes a driver monitoring feature
+that alerts when it detects driver distraction.
 
 However, even with an attentive driver, we must make further efforts for the system to be
 safe. We repeat, **driver alertness is necessary, but not sufficient, for openpilot to be

docs/ext/glossary.py

@@ -8,7 +8,7 @@ from markdown.extensions import Extension
 from markdown.preprocessors import Preprocessor
 from markdown.treeprocessors import Treeprocessor
 
-from zensical.extensions.links import LinksProcessor
+from zensical.extensions.links import LinksTreeprocessor
 
 GlossaryTerm = tuple[str, re.Pattern[str], str]
 
@@ -78,7 +78,7 @@ class GlossaryTreeprocessor(Treeprocessor):
   def run(self, root: ET.Element) -> None:
     at = self.md.treeprocessors.get_index_for_name("zrelpath")
     processor = self.md.treeprocessors[at]
-    if not isinstance(processor, LinksProcessor):
+    if not isinstance(processor, LinksTreeprocessor):
       raise TypeError("Links processor not registered")
     if processor.path == GLOSSARY_PAGE:
       return

docs/how-to/turn-the-speed-blue.md

@@ -20,7 +20,7 @@ source .venv/bin/activate
 
 Then, compile openpilot:
 ```bash
-scons -j$(nproc)
+scons
 ```
 
 ## 2. Run replay

launch_env.sh

@@ -16,7 +16,7 @@ export VECLIB_MAXIMUM_THREADS=1
 export QCOM_PRIORITY=12
 
 if [ -z "$AGNOS_VERSION" ]; then
-  export AGNOS_VERSION="17.2"
+  export AGNOS_VERSION="18.4"
 fi
 
 export STAGING_ROOT="/data/safe_staging"

openpilot/third_party

@@ -1 +0,0 @@
-../third_party

pyproject.toml

@@ -20,14 +20,17 @@ dependencies = [
   # core
   "cffi",
   "scons",
-  "pycapnp",
+  "pycapnp==2.1.0",  # 2.2 introduces a memory leak due to cyclic references
   "Cython",
   "setuptools",
   "numpy >=2.0",
 
   # vendored native dependencies
   "bzip2 @ git+https://github.com/commaai/dependencies.git@release-bzip2#subdirectory=bzip2",
+  "bootstrap-icons @ git+https://github.com/commaai/dependencies.git@release-bootstrap-icons#subdirectory=bootstrap-icons",
   "capnproto @ git+https://github.com/commaai/dependencies.git@release-capnproto#subdirectory=capnproto",
+  "catch2 @ git+https://github.com/commaai/dependencies.git@release-catch2#subdirectory=catch2",
+  "acados @ git+https://github.com/commaai/dependencies.git@release-acados#subdirectory=acados",
   "eigen @ git+https://github.com/commaai/dependencies.git@release-eigen#subdirectory=eigen",
   "ffmpeg @ git+https://github.com/commaai/dependencies.git@release-ffmpeg#subdirectory=ffmpeg",
   "libjpeg @ git+https://github.com/commaai/dependencies.git@release-libjpeg#subdirectory=libjpeg",
@@ -36,8 +39,10 @@ dependencies = [
   "ncurses @ git+https://github.com/commaai/dependencies.git@release-ncurses#subdirectory=ncurses",
   "zeromq @ git+https://github.com/commaai/dependencies.git@release-zeromq#subdirectory=zeromq",
   "libusb @ git+https://github.com/commaai/dependencies.git@release-libusb#subdirectory=libusb",
+  "json11 @ git+https://github.com/commaai/dependencies.git@release-json11#subdirectory=json11",
   "git-lfs @ git+https://github.com/commaai/dependencies.git@release-git-lfs#subdirectory=git-lfs",
   "gcc-arm-none-eabi @ git+https://github.com/commaai/dependencies.git@release-gcc-arm-none-eabi#subdirectory=gcc-arm-none-eabi",
+  "xvfb @ git+https://github.com/commaai/dependencies.git@release-xvfb#subdirectory=xvfb",
 
   # body / webrtcd
   "av",
@@ -58,9 +63,6 @@ dependencies = [
   "json-rpc",
   "websocket_client",
 
-  # acados deps
-  "casadi >=3.6.6",  # 3.12 fixed in 3.6.6
-
   # joystickd
   "inputs",
 
@@ -73,7 +75,7 @@ dependencies = [
   "zstandard",
 
   # ui
-  "raylib > 5.5.0.3",
+  "raylib @ git+https://github.com/commaai/dependencies.git@release-raylib#subdirectory=raylib",
   "qrcode",
   "jeepney",
   "pillow",
@@ -94,7 +96,6 @@ testing = [
   "pytest-subtests",
   # https://github.com/pytest-dev/pytest-xdist/pull/1229
   "pytest-xdist @ git+https://github.com/sshane/pytest-xdist@2b4372bd62699fb412c4fe2f95bf9f01bd2018da",
-  "pytest-asyncio",
   "pytest-mock",
   "ruff",
   "codespell",
@@ -107,7 +108,7 @@ dev = [
 ]
 
 tools = [
-  "imgui  @ git+https://github.com/commaai/dependencies.git@release-imgui#subdirectory=imgui",
+  "imgui @ git+https://github.com/commaai/dependencies.git@release-imgui#subdirectory=imgui",
   "metadrive-simulator @ git+https://github.com/commaai/metadrive.git@minimal ; (platform_machine != 'aarch64')",
 ]
 
@@ -126,15 +127,17 @@ allow-direct-references = true
 
 [tool.pytest.ini_options]
 minversion = "6.0"
-addopts = "--ignore=openpilot/ --ignore=opendbc/ --ignore=panda/ --ignore=rednose_repo/ --ignore=tinygrad_repo/ --ignore=teleoprtc_repo/ --ignore=msgq/  -Werror --strict-config --strict-markers --durations=10 -n auto --dist=loadgroup"
+addopts = "--ignore=openpilot/ --ignore=opendbc/ --ignore=panda/ --ignore=rednose_repo/ --ignore=tinygrad_repo/ --ignore=teleoprtc_repo/ --ignore=msgq/ -Werror --strict-config --strict-markers --durations=10 -n auto --dist=loadgroup"
 cpp_files = "test_*"
 cpp_harness = "selfdrive/test/cpp_harness.py"
 python_files = "test_*.py"
-asyncio_default_fixture_loop_scope = "function"
 markers = [
   "slow: tests that take awhile to run and can be skipped with -m 'not slow'",
   "tici: tests that are only meant to run on the C3/C3X",
-  "skip_tici_setup: mark test to skip tici setup fixture"
+  "skip_tici_setup: mark test to skip tici setup fixture",
+  "nocapture: don't capture test output",
+  "shared_download_cache: share download cache between tests",
+  "xdist_group_class_property: group tests by a property of the class that contains them",
 ]
 testpaths = [
   "common",
@@ -175,9 +178,7 @@ lint.ignore = [
   "UP045", "UP007",  # these don't play nice with raylib atm
 ]
 line-length = 160
-target-version ="py311"
 exclude = [
-  "body",
   "cereal",
   "panda",
   "opendbc",
@@ -186,7 +187,7 @@ exclude = [
   "tinygrad_repo",
   "teleoprtc",
   "teleoprtc_repo",
-  "third_party",
+  "third_party/copyparty",
   "*.ipynb",
   "generated",
 ]
@@ -196,7 +197,6 @@ lint.flake8-implicit-str-concat.allow-multiline = false
 "selfdrive".msg = "Use openpilot.selfdrive"
 "common".msg = "Use openpilot.common"
 "system".msg = "Use openpilot.system"
-"third_party".msg = "Use openpilot.third_party"
 "tools".msg = "Use openpilot.tools"
 "pytest.main".msg = "pytest.main requires special handling that is easy to mess up!"
 "unittest".msg = "Use pytest"
@@ -214,7 +214,6 @@ quote-style = "preserve"
 
 [tool.ty.src]
 exclude = [
-  "cereal/",
   "msgq/",
   "msgq_repo/",
   "opendbc/",
@@ -230,27 +229,16 @@ exclude = [
 ]
 
 [tool.ty.rules]
-# Ignore unresolved imports for Cython-compiled modules (.pyx)
-unresolved-import = "ignore"
-# Ignore unresolved attributes - many from capnp and Cython modules
-unresolved-attribute = "ignore"
-# Ignore invalid method overrides - signature variance issues
-invalid-method-override = "ignore"
-# Ignore possibly-missing-attribute - too many false positives
-possibly-missing-attribute = "ignore"
-# Ignore invalid assignment - often intentional monkey-patching
-invalid-assignment = "ignore"
-# Ignore no-matching-overload - numpy/ctypes overload matching issues
-no-matching-overload = "ignore"
-# Ignore invalid-argument-type - many false positives from raylib, ctypes, numpy
-invalid-argument-type = "ignore"
-# Ignore call-non-callable - false positives from dynamic types
-call-non-callable = "ignore"
-# Ignore unsupported-operator - false positives from dynamic types
-unsupported-operator = "ignore"
-# Ignore not-subscriptable - false positives from dynamic types
-not-subscriptable = "ignore"
-# not-iterable errors are now fixed
+unresolved-import = "ignore"  # Cython-compiled modules (.pyx)
+unresolved-attribute = "ignore"  # many from capnp and Cython modules
+invalid-method-override = "ignore"  # signature variance issues
+possibly-missing-attribute = "ignore"  # too many false positives
+invalid-assignment = "ignore"  # often intentional monkey-patching
+no-matching-overload = "ignore"  # numpy/ctypes overload matching issues
+invalid-argument-type = "ignore"  # many false positives from raylib, ctypes, numpy
+call-non-callable = "ignore"  # false positives from dynamic types
+unsupported-operator = "ignore" # false positives from dynamic types
+not-subscriptable = "ignore" # false positives from dynamic types
 
 [tool.uv]
 python-preference = "only-managed"

release/build_release.sh

@@ -16,6 +16,8 @@ if [ -z "$RELEASE_BRANCH" ]; then
   exit 1
 fi
 
+BUILD_BRANCH=release-mici-staging
+
 
 # set git identity
 source $DIR/identity.sh
@@ -26,7 +28,7 @@ mkdir -p $BUILD_DIR
 cd $BUILD_DIR
 git init
 git remote add origin git@github.com:commaai/openpilot.git
-git checkout --orphan $RELEASE_BRANCH
+git checkout --orphan $BUILD_BRANCH
 
 # do the files copy
 echo "[-] copying files T=$SECONDS"
@@ -46,14 +48,14 @@ git commit -a -m "openpilot v$VERSION release"
 
 # Build
 export PYTHONPATH="$BUILD_DIR"
-scons -j$(nproc) --minimal
+scons
 
 if [ -z "$PANDA_DEBUG_BUILD" ]; then
   # release panda fw
-  CERT=/data/pandaextra/certs/release RELEASE=1 scons -j$(nproc) panda/
+  CERT=/data/pandaextra/certs/release RELEASE=1 scons panda/
 else
   # build with ALLOW_DEBUG=1 to enable features like experimental longitudinal
-  scons -j$(nproc) panda/
+  scons panda/
 fi
 
 # Ensure no submodules in release
@@ -72,8 +74,8 @@ find . -name '*.pyc' -delete
 find . -name 'moc_*' -delete
 find . -name '__pycache__' -delete
 rm -rf .sconsign.dblite Jenkinsfile release/
-rm -f selfdrive/modeld/models/*.onnx
-rm -f sunnypilot/modeld*/models/*.onnx
+rm -f selfdrive/modeld/models/*.onnx*
+rm -f sunnypilot/modeld*/models/*.onnx*
 
 find third_party/ -name '*x86*' -exec rm -r {} +
 find third_party/ -name '*Darwin*' -exec rm -r {} +
@@ -94,9 +96,11 @@ cd $BUILD_DIR
 RELEASE=1 pytest -n0 -s selfdrive/test/test_onroad.py
 #pytest selfdrive/car/tests/test_car_interfaces.py
 
-if [ ! -z "$RELEASE_BRANCH" ]; then
-  echo "[-] pushing release T=$SECONDS"
-  git push -f origin $RELEASE_BRANCH:$RELEASE_BRANCH
-fi
+echo "[-] pushing release T=$SECONDS"
+REFS=()
+for branch in ${RELEASE_BRANCH//,/ }; do
+  REFS+=("$BUILD_BRANCH:$branch")
+done
+git push -f origin "${REFS[@]}"
 
 echo "[-] done T=$SECONDS"

release/build_stripped.sh

@@ -45,6 +45,8 @@ cd $TARGET_DIR
 rm -rf .git/modules/
 rm -f panda/board/obj/panda.bin.signed
 
+find selfdrive/modeld/models -name '*.onnx' -size +95M -exec ./common/file_chunker.py {} \;
+
 # include source commit hash and build date in commit
 GIT_HASH=$(git --git-dir=$SOURCE_DIR/.git rev-parse HEAD)
 GIT_COMMIT_DATE=$(git --git-dir=$SOURCE_DIR/.git show --no-patch --format='%ct %ci' HEAD)

release/ci/model_generator.py

@@ -1,3 +1,10 @@
+"""
+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 pickle
 import sys
@@ -32,6 +39,9 @@ OPTIONAL_OUTPUT_KEYS = frozenset({
 def validate_model_outputs(metadata_paths: list[Path]) -> None:
   combined_keys: set[str] = set()
   for path in metadata_paths:
+    if path.stat().st_size == 0:
+      print(f"skipping empty metadata: {path}")
+      continue
     with open(path, "rb") as f:
       metadata = pickle.load(f)
     combined_keys.update(metadata.get("output_slices", {}).keys())
@@ -78,38 +88,65 @@ def create_short_name(full_name):
   return result[:8]
 
 
-def generate_metadata(model_path: Path, output_dir: Path, short_name: str):
-  model_path = model_path
-  output_path = output_dir
+def _read_pkl_bytes(pkl_path: Path) -> bytes:
+  manifest = Path(f"{pkl_path}.chunkmanifest")
+  if manifest.exists():
+    num_chunks = int(manifest.read_text().strip())
+    parts = []
+    for i in range(num_chunks):
+      chunk = Path(f"{pkl_path}.chunk{i + 1:02d}of{num_chunks:02d}")
+      parts.append(chunk.read_bytes())
+    return b''.join(parts)
+  return pkl_path.read_bytes()
+
+
+def _find_driving_pkl(output_path: Path) -> Path | None:
+  for pattern in ('driving_tinygrad.pkl', 'driving_*_tinygrad.pkl'):
+    matches = sorted(output_path.glob(pattern))
+    if matches:
+      return matches[0]
+  for pattern in ('driving_tinygrad.pkl.chunkmanifest', 'driving_*_tinygrad.pkl.chunkmanifest'):
+    matches = sorted(output_path.glob(pattern))
+    if matches:
+      return Path(str(matches[0]).removesuffix('.chunkmanifest'))
+  return None
+
+
+def _rename_pkl_with_chunks(old_pkl: Path, new_pkl: Path) -> Path:
+  manifest = Path(f"{old_pkl}.chunkmanifest")
+  if manifest.exists():
+    for f in sorted(old_pkl.parent.glob(f"{old_pkl.name}.chunk*")):
+      f.rename(old_pkl.parent / f.name.replace(old_pkl.name, new_pkl.name, 1))
+    return new_pkl
+  return old_pkl.rename(new_pkl)
+
+
+def generate_metadata(model_path: Path, output_dir: Path, short_name: str, driving_pkl: Path):
   base = model_path.stem
+  metadata_file = output_dir / f"{base}_metadata.pkl"
 
-  # Define output files for tinygrad and metadata
-  tinygrad_file = output_path / f"{base}_tinygrad.pkl"
-  metadata_file = output_path / f"{base}_metadata.pkl"
+  if short_name:
+    renamed_meta = output_dir / f"{base}_{short_name.lower()}_metadata.pkl"
+    if metadata_file.exists() and not renamed_meta.exists():
+      metadata_file = metadata_file.rename(renamed_meta)
+    elif renamed_meta.exists():
+      metadata_file = renamed_meta
 
-  if not tinygrad_file.exists() or not metadata_file.exists():
-    print(f"Error: Missing files for model {base} ({tinygrad_file} or {metadata_file})", file=sys.stderr)
+  if not metadata_file.exists():
+    print(f"Warning: Missing metadata for {base} ({metadata_file}), skipping", file=sys.stderr)
     return
 
-  # Calculate the sha256 hashes
-  with open(tinygrad_file, 'rb') as f:
-    tinygrad_hash = hashlib.sha256(f.read()).hexdigest()
+  tinygrad_hash = hashlib.sha256(_read_pkl_bytes(driving_pkl)).hexdigest()
 
   with open(metadata_file, 'rb') as f:
     metadata_hash = hashlib.sha256(f.read()).hexdigest()
 
-  # Rename the files if a custom file name is provided
-  if short_name:
-    tinygrad_file = tinygrad_file.rename(output_path / f"{base}_{short_name.lower()}_tinygrad.pkl")
-    metadata_file = metadata_file.rename(output_path / f"{base}_{short_name.lower()}_metadata.pkl")
-
-  # Build the metadata structure
   model_type = "offPolicy" if "off_policy" in base else "onPolicy" if "on_policy" in base else base.split("_")[-1]
 
-  model_metadata = {
+  return {
     "type": model_type,
     "artifact": {
-      "file_name": tinygrad_file.name,
+      "file_name": driving_pkl.name,
       "download_uri": {
         "url": "https://gitlab.com/sunnypilot/public/docs.sunnypilot.ai/-/raw/main/",
         "sha256": tinygrad_hash
@@ -124,9 +161,6 @@ def generate_metadata(model_path: Path, output_dir: Path, short_name: str):
     }
   }
 
-  # Return model metadata
-  return model_metadata
-
 
 def create_metadata_json(models: list, output_dir: Path, custom_name=None, short_name=None, is_20hz=False, upstream_branch="unknown"):
   metadata_json = {
@@ -181,14 +215,28 @@ if __name__ == "__main__":
 
   _output_dir = Path(args.output_dir)
   _output_dir.mkdir(exist_ok=True, parents=True)
+  _short_name = create_short_name(args.custom_name) if args.custom_name else None
+
+  _driving_pkl = _find_driving_pkl(_output_dir)
+  if not _driving_pkl:
+    print(f"No driving_tinygrad.pkl found in {_output_dir}", file=sys.stderr)
+    sys.exit(1)
+
+  if _short_name:
+    new_pkl = _output_dir / f"driving_{_short_name.lower()}_tinygrad.pkl"
+    if not new_pkl.exists():
+      _driving_pkl = _rename_pkl_with_chunks(_driving_pkl, new_pkl)
+    else:
+      _driving_pkl = new_pkl
+
   _models = []
 
   for _model_path in model_paths:
-    _model_metadata = generate_metadata(Path(_model_path), _output_dir, create_short_name(args.custom_name))
+    _model_metadata = generate_metadata(Path(_model_path), _output_dir, _short_name, _driving_pkl)
     if _model_metadata:
       _models.append(_model_metadata)
 
   if _models:
-    create_metadata_json(_models, _output_dir, args.custom_name, create_short_name(args.custom_name), args.is_20hz, args.upstream_branch)
+    create_metadata_json(_models, _output_dir, args.custom_name, _short_name, args.is_20hz, args.upstream_branch)
   else:
     print("No models processed.", file=sys.stderr)

release/pack.py

@@ -13,7 +13,7 @@ from openpilot.common.basedir import BASEDIR
 
 DIRS = ['cereal', 'openpilot']
 EXTS = ['.png', '.py', '.ttf', '.capnp', '.json', '.fnt', '.mo', '.po']
-EXCLUDE = ['selfdrive/assets/training', 'third_party/raylib/raylib_repo/examples']
+EXCLUDE = ['selfdrive/assets/training']
 INTERPRETER = '/usr/bin/env python3'
 
 

scripts/lint/check_raylib_includes.sh

@@ -1,10 +0,0 @@
-#!/usr/bin/env bash
-
-FAIL=0
-
-if grep -n '#include "third_party/raylib/include/raylib\.h"' $@ | grep -v '^system/ui/raylib/raylib\.h'; then
-  echo -e "Bad raylib include found! Use '#include \"system/ui/raylib/raylib.h\"' instead\n"
-  FAIL=1
-fi
-
-exit $FAIL

scripts/lint/lint.sh

@@ -14,7 +14,7 @@ cd $ROOT
 FAILED=0
 
 IGNORED_FILES="uv\.lock|docs\/CARS.md|LICENSE\.md"
-IGNORED_DIRS="^third_party.*|^msgq.*|^msgq_repo.*|^opendbc.*|^opendbc_repo.*|^cereal.*|^panda.*|^rednose.*|^rednose_repo.*|^tinygrad.*|^tinygrad_repo.*|^teleoprtc.*|^teleoprtc_repo.*"
+IGNORED_DIRS="^msgq.*|^msgq_repo.*|^opendbc.*|^opendbc_repo.*|^cereal.*|^panda.*|^rednose.*|^rednose_repo.*|^tinygrad.*|^tinygrad_repo.*|^teleoprtc.*|^teleoprtc_repo.*|^third_party.*"
 
 function run() {
   shopt -s extglob

scripts/reporter.py

@@ -34,6 +34,11 @@ if __name__ == "__main__":
 
   for f in glob.glob(BASEDIR + MODEL_PATH + "/*.onnx"):
     fn = os.path.basename(f)
-    master = get_checkpoint(MASTER_PATH + MODEL_PATH + fn)
+    master_path = MASTER_PATH + MODEL_PATH + fn
+    if os.path.exists(master_path):
+      master = get_checkpoint(master_path)
+      master_col = f"[{master}](https://reporter.comma.life/experiment/{master})"
+    else:
+      master_col = "N/A (new model)"
     pr = get_checkpoint(BASEDIR + MODEL_PATH + fn)
-    print("|", fn, "|", f"[{master}](https://reporterv2.comma.life/{master})", "|", f"[{pr}](https://reporterv2.comma.life/{pr})", "|")
+    print("|", fn, "|", master_col, "|", f"[{pr}](https://reporter.comma.life/experiment/{pr})", "|")

selfdrive/assets/assets.qrc

@@ -1,6 +1,6 @@
 <!DOCTYPE RCC><RCC version="1.0">
 <qresource>
-  <file alias="bootstrap-icons.svg">../../third_party/bootstrap/bootstrap-icons.svg</file>
+  <file alias="bootstrap-icons.svg">@BOOTSTRAP_ICONS_SVG@</file>
   <file>images/button_continue_triangle.svg</file>
   <file>icons/circled_check.svg</file>
   <file>icons/circled_slash.svg</file>

selfdrive/assets/fonts/process.py

@@ -37,10 +37,10 @@ def _char_sets():
   return tuple(sorted(ord(c) for c in base)), tuple(sorted(ord(c) for c in unifont))
 
 
-def _glyph_metrics(glyphs, rects, codepoints):
+def _glyph_metrics(glyphs, rects, glyph_count: int):
   entries = []
   min_offset_y, max_extent = None, 0
-  for idx, codepoint in enumerate(codepoints):
+  for idx in range(glyph_count):
     glyph = glyphs[idx]
     rect = rects[idx]
     width = int(round(rect.width))
@@ -49,7 +49,7 @@ def _glyph_metrics(glyphs, rects, codepoints):
     min_offset_y = offset_y if min_offset_y is None else min(min_offset_y, offset_y)
     max_extent = max(max_extent, offset_y + height)
     entries.append({
-      "id": codepoint,
+      "id": glyph.value,
       "x": int(round(rect.x)),
       "y": int(round(rect.y)),
       "width": width,
@@ -97,19 +97,23 @@ def _process_font(font_path: Path, codepoints: tuple[int, ...]):
   file_buf = rl.ffi.new("unsigned char[]", data)
   cp_buffer = rl.ffi.new("int[]", codepoints)
   cp_ptr = rl.ffi.cast("int *", cp_buffer)
-  glyphs = rl.load_font_data(rl.ffi.cast("unsigned char *", file_buf), len(data), font_size, cp_ptr, len(codepoints), rl.FontType.FONT_DEFAULT)
+  glyph_count = rl.ffi.new("int *", len(codepoints))
+  glyphs = rl.load_font_data(
+    rl.ffi.cast("unsigned char *", file_buf), len(data), font_size, cp_ptr, len(codepoints),
+    rl.FontType.FONT_DEFAULT, glyph_count
+  )
   if glyphs == rl.ffi.NULL:
     raise RuntimeError("raylib failed to load font data")
 
   rects_ptr = rl.ffi.new("Rectangle **")
-  image = rl.gen_image_font_atlas(glyphs, rects_ptr, len(codepoints), font_size, GLYPH_PADDING, 0)
+  image = rl.gen_image_font_atlas(glyphs, rects_ptr, glyph_count[0], font_size, GLYPH_PADDING, 0)
   if image.width == 0 or image.height == 0:
     raise RuntimeError("raylib returned an empty atlas")
 
   rects = rects_ptr[0]
   atlas_name = f"{font_path.stem}.png"
   atlas_path = FONT_DIR / atlas_name
-  entries, line_height, base = _glyph_metrics(glyphs, rects, codepoints)
+  entries, line_height, base = _glyph_metrics(glyphs, rects, glyph_count[0])
 
   if not rl.export_image(image, atlas_path.as_posix()):
     raise RuntimeError("Failed to export atlas image")

selfdrive/assets/icons_mici/egpu.png

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec3dcf64cbc34251d8423cb8b3b31d743e93d14002dec43c389a857cb7e8eb17
+size 10875

selfdrive/assets/icons_mici/egpu_gray.png

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7409c53d7c72681c24982fd83b56ce70f80797c9c0f936d9296a5c18557ac472
+size 7279

selfdrive/assets/prep-svg.sh

@@ -3,7 +3,7 @@ set -e
 
 DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null && pwd )"
 ICONS_DIR="$DIR/icons"
-BOOTSTRAP_SVG="$DIR/../../third_party/bootstrap/bootstrap-icons.svg"
+BOOTSTRAP_SVG="$(python3 -c 'import bootstrap_icons; print(bootstrap_icons.SVG_PATH)')"
 
 ICON_IDS=(
   arrow-right

selfdrive/car/car_specific.py

@@ -73,7 +73,7 @@ class CarSpecificEvents:
     elif self.CP.brand == 'gm':
       # Enabling at a standstill with brake is allowed
       # TODO: verify 17 Volt can enable for the first time at a stop and allow for all GMs
-      if CS.vEgo < self.CP.minEnableSpeed and not (CS.standstill and CS.brake >= 20 and
+      if CS.vEgo < self.CP.minEnableSpeed and not (CS.standstill and CS.brakePressed and
                                                    self.CP.networkLocation == NetworkLocation.fwdCamera):
         events.add(EventName.belowEngageSpeed)
       if CS.cruiseState.standstill:

selfdrive/car/card.py

@@ -10,7 +10,7 @@ from cereal import car, log, custom
 from openpilot.common.params import Params
 from openpilot.common.realtime import config_realtime_process, Priority, Ratekeeper
 from openpilot.common.swaglog import cloudlog, ForwardingHandler
-
+from opendbc.safety import ALTERNATIVE_EXPERIENCE
 from opendbc.car import DT_CTRL, structs
 from opendbc.car.can_definitions import CanData, CanRecvCallable, CanSendCallable
 from opendbc.car.carlog import carlog
@@ -37,7 +37,7 @@ def obd_callback(params: Params) -> ObdCallback:
     if params.get_bool("ObdMultiplexingEnabled") != obd_multiplexing:
       cloudlog.warning(f"Setting OBD multiplexing to {obd_multiplexing}")
       params.remove("ObdMultiplexingChanged")
-      params.put_bool("ObdMultiplexingEnabled", obd_multiplexing)
+      params.put_bool("ObdMultiplexingEnabled", obd_multiplexing, block=True)
       params.get_bool("ObdMultiplexingChanged", block=True)
       cloudlog.warning("OBD multiplexing set successfully")
   return set_obd_multiplexing
@@ -86,7 +86,7 @@ class Car:
 
     self.can_callbacks = can_comm_callbacks(self.can_sock, self.pm.sock['sendcan'])
 
-    is_release = self.params.get_bool("IsReleaseBranch")
+    is_release = False  # self.params.get_bool("IsReleaseBranch")
     is_release_sp = self.params.get_bool("IsReleaseSpBranch")
 
     if CI is None:
@@ -116,12 +116,18 @@ class Car:
       self.CP_SP = self.CI.CP_SP
 
       # continue onto next fingerprinting step in pandad
-      self.params.put_bool("FirmwareQueryDone", True)
+      self.params.put_bool("FirmwareQueryDone", True, block=True)
     else:
       self.CI, self.CP, self.CP_SP = CI, CI.CP, CI.CP_SP
       self.RI = RI
 
+    # set alternative experiences from parameters
+    sp_toyota_auto_brake_hold = self.params.get_bool("ToyotaAutoHold")
     self.CP.alternativeExperience = 0
+    if sp_toyota_auto_brake_hold:
+      self.CP.alternativeExperience |= ALTERNATIVE_EXPERIENCE.ALLOW_AEB
+
+
     # mads
     set_alternative_experience(self.CP, self.CP_SP, self.params)
     set_car_specific_params(self.CP, self.CP_SP, self.params)
@@ -143,7 +149,7 @@ class Car:
         with open("/cache/params/SecOCKey") as f:
           user_key = f.readline().strip()
           if len(user_key) == 32:
-            self.params.put("SecOCKey", user_key)
+            self.params.put("SecOCKey", user_key, block=True)
       except Exception:
         pass
 
@@ -161,21 +167,21 @@ class Car:
     # Write previous route's CarParams
     prev_cp = self.params.get("CarParamsPersistent")
     if prev_cp is not None:
-      self.params.put("CarParamsPrevRoute", prev_cp)
+      self.params.put("CarParamsPrevRoute", prev_cp, block=True)
 
     # Write CarParams for controls and radard
     cp_bytes = self.CP.to_bytes()
-    self.params.put("CarParams", cp_bytes)
-    self.params.put_nonblocking("CarParamsCache", cp_bytes)
-    self.params.put_nonblocking("CarParamsPersistent", cp_bytes)
+    self.params.put("CarParams", cp_bytes, block=True)
+    self.params.put("CarParamsCache", cp_bytes)
+    self.params.put("CarParamsPersistent", cp_bytes)
 
     # Write CarParamsSP for controls
     # convert to pycapnp representation for caching and logging
     self.CP_SP_capnp = convert_to_capnp(self.CP_SP)
     cp_sp_bytes = self.CP_SP_capnp.to_bytes()
-    self.params.put("CarParamsSP", cp_sp_bytes)
-    self.params.put_nonblocking("CarParamsSPCache", cp_sp_bytes)
-    self.params.put_nonblocking("CarParamsSPPersistent", cp_sp_bytes)
+    self.params.put("CarParamsSP", cp_sp_bytes, block=True)
+    self.params.put("CarParamsSPCache", cp_sp_bytes)
+    self.params.put("CarParamsSPPersistent", cp_sp_bytes)
 
     self.v_cruise_helper = VCruiseHelper(self.CP, self.CP_SP)
 
@@ -274,7 +280,7 @@ class Car:
       # TODO: this can make us miss at least a few cycles when doing an ECU knockout
       self.CI.init(self.CP, self.CP_SP, *self.can_callbacks)
       # signal pandad to switch to car safety mode
-      self.params.put_bool_nonblocking("ControlsReady", True)
+      self.params.put_bool("ControlsReady", True)
 
     if self.sm.all_alive(['carControl']):
       # send car controls over can

selfdrive/car/tests/test_cruise_speed.py

@@ -94,7 +94,7 @@ class TestVCruiseHelper:
         self.enable(V_CRUISE_INITIAL * CV.KPH_TO_MS, False, False)
 
       # Expected diff on enabling. Speed should not change on falling edge of pressed
-      assert not pressed == self.v_cruise_helper.v_cruise_kph == self.v_cruise_helper.v_cruise_kph_last
+      assert (not pressed) == (self.v_cruise_helper.v_cruise_kph == self.v_cruise_helper.v_cruise_kph_last)
 
   def test_resume_in_standstill(self):
     """

selfdrive/car/tests/test_models.py

@@ -13,7 +13,7 @@ from opendbc.car import DT_CTRL, gen_empty_fingerprint, structs
 from opendbc.car.can_definitions import CanData
 from opendbc.car.car_helpers import FRAME_FINGERPRINT, interfaces
 from opendbc.car.fingerprints import MIGRATION
-from opendbc.car.honda.values import CAR as HONDA, HondaFlags
+from opendbc.car.honda.values import HondaFlags
 from opendbc.car.structs import car
 from opendbc.car.tests.routes import non_tested_cars, routes, CarTestRoute
 from opendbc.car.values import Platform, PLATFORMS
@@ -28,6 +28,14 @@ from openpilot.tools.lib.route import SegmentName
 SafetyModel = car.CarParams.SafetyModel
 SteerControlType = structs.CarParams.SteerControlType
 
+# panda safety stores angle_meas in brand-specific CAN units (angle_deg_to_can in opendbc/safety/modes/*.h).
+ANGLE_DEG_TO_CAN = {
+  "tesla": -10,
+  "toyota": 17.452007,
+  "nissan": 100,
+  "psa": 10,
+}
+
 NUM_JOBS = int(os.environ.get("NUM_JOBS", "1"))
 JOB_ID = int(os.environ.get("JOB_ID", "0"))
 INTERNAL_SEG_LIST = os.environ.get("INTERNAL_SEG_LIST", "")
@@ -350,13 +358,7 @@ class TestCarModelBase(unittest.TestCase):
         self.assertEqual(CS.gasPressed, self.safety.get_gas_pressed_prev())
 
       if self.safety.get_brake_pressed_prev() != prev_panda_brake:
-        # TODO: remove this exception once this mismatch is resolved
-        brake_pressed = CS.brakePressed
-        if CS.brakePressed and not self.safety.get_brake_pressed_prev():
-          if self.CP.carFingerprint in (HONDA.HONDA_PILOT, HONDA.HONDA_RIDGELINE) and CS.brake > 0.05:
-            brake_pressed = False
-
-        self.assertEqual(brake_pressed, self.safety.get_brake_pressed_prev())
+        self.assertEqual(CS.brakePressed, self.safety.get_brake_pressed_prev())
 
       if self.safety.get_regen_braking_prev() != prev_panda_regen_braking:
         self.assertEqual(CS.regenBraking, self.safety.get_regen_braking_prev())
@@ -433,12 +435,14 @@ class TestCarModelBase(unittest.TestCase):
         checks['vEgoRaw'] += (v_ego_raw > (self.safety.get_vehicle_speed_max() + 1e-3) or
                               v_ego_raw < (self.safety.get_vehicle_speed_min() - 1e-3))
 
-      # TODO: remove this exception once this mismatch is resolved
-      brake_pressed = CS.brakePressed
-      if CS.brakePressed and not self.safety.get_brake_pressed_prev():
-        if self.CP.carFingerprint in (HONDA.HONDA_PILOT, HONDA.HONDA_RIDGELINE) and CS.brake > 0.05:
-          brake_pressed = False
-      checks['brakePressed'] += brake_pressed != self.safety.get_brake_pressed_prev()
+      # check steering angle for angle control cars (panda stores angle_meas in CAN units)
+      # ford excluded since it tracks curvature, not steering angle
+      if self.CP.steerControlType == SteerControlType.angle and not self.CP.notCar and self.CP.brand != "ford":
+        angle_can = (CS.steeringAngleDeg + CS.steeringAngleOffsetDeg) * ANGLE_DEG_TO_CAN[self.CP.brand]
+        checks['steeringAngleDeg'] += (angle_can > (self.safety.get_angle_meas_max() + 1) or
+                                       angle_can < (self.safety.get_angle_meas_min() - 1))
+
+      checks['brakePressed'] += CS.brakePressed != self.safety.get_brake_pressed_prev()
       checks['regenBraking'] += CS.regenBraking != self.safety.get_regen_braking_prev()
       checks['steeringDisengage'] += CS.steeringDisengage != self.safety.get_steering_disengage_prev()
 

selfdrive/controls/controlsd.py

@@ -212,7 +212,7 @@ class Controls(ControlsExt):
     cs.upAccelCmd = float(self.LoC.pid.p)
     cs.uiAccelCmd = float(self.LoC.pid.i)
     cs.ufAccelCmd = float(self.LoC.pid.f)
-    cs.forceDecel = bool((self.sm['driverMonitoringState'].awarenessStatus < 0.) or
+    cs.forceDecel = bool((self.sm['driverMonitoringState'].alertLevel == log.DriverMonitoringState.AlertLevel.three) or
                          (self.sm['selfdriveState'].state == State.softDisabling))
 
     lat_tuning = self.CP.lateralTuning.which()

selfdrive/controls/lib/desire_helper.py

@@ -56,7 +56,7 @@ class DesireHelper:
   def get_lane_change_direction(CS):
     return LaneChangeDirection.left if CS.leftBlinker else LaneChangeDirection.right
 
-  def update(self, carstate, lateral_active, lane_change_prob):
+  def update(self, carstate, lateral_active, lane_change_prob, left_edge_detected=False, right_edge_detected=False):
     self.alc.update_params()
     self.lane_turn_controller.update_params()
     v_ego = carstate.vEgo
@@ -88,8 +88,8 @@ class DesireHelper:
                          ((carstate.steeringTorque > 0 and self.lane_change_direction == LaneChangeDirection.left) or
                           (carstate.steeringTorque < 0 and self.lane_change_direction == LaneChangeDirection.right))
 
-        blindspot_detected = ((carstate.leftBlindspot and self.lane_change_direction == LaneChangeDirection.left) or
-                              (carstate.rightBlindspot and self.lane_change_direction == LaneChangeDirection.right))
+        blindspot_detected = (((carstate.leftBlindspot or left_edge_detected) and self.lane_change_direction == LaneChangeDirection.left) or
+                              ((carstate.rightBlindspot or right_edge_detected) and self.lane_change_direction == LaneChangeDirection.right))
 
         self.alc.update_lane_change(blindspot_detected, carstate.brakePressed)
 

selfdrive/controls/lib/drive_helpers.py

@@ -8,6 +8,7 @@ CAR_ROTATION_RADIUS = 0.0
 # This is a turn radius smaller than most cars can achieve
 MAX_CURVATURE = 0.2
 MAX_VEL_ERR = 5.0  # m/s
+MIN_STABLE_DELAY = 0.3
 
 # EU guidelines
 MAX_LATERAL_JERK = 5.0  # m/s^3
@@ -43,7 +44,10 @@ def get_accel_from_plan(speeds, accels, t_idxs, action_t=DT_MDL, vEgoStopping=0.
   if len(speeds) == len(t_idxs):
     v_now = speeds[0]
     a_now = accels[0]
-    v_target = np.interp(action_t, t_idxs, speeds)
+    if action_t < MIN_STABLE_DELAY:
+      v_target = v_now + (action_t / MIN_STABLE_DELAY) * (np.interp(MIN_STABLE_DELAY, t_idxs, speeds) - v_now)
+    else:
+      v_target = np.interp(action_t, t_idxs, speeds)
     a_target = 2 * (v_target - v_now) / (action_t) - a_now
   else:
     v_now = 0.0
@@ -58,6 +62,9 @@ def curv_from_psis(psi_target, psi_rate, vego, action_t):
   return 2*curv_from_psi - psi_rate / vego
 
 def get_curvature_from_plan(yaws, yaw_rates, t_idxs, vego, action_t):
-  psi_target = np.interp(action_t, t_idxs, yaws)
+  if action_t < MIN_STABLE_DELAY:
+    psi_target = (action_t / MIN_STABLE_DELAY) * np.interp(MIN_STABLE_DELAY, t_idxs, yaws)
+  else:
+    psi_target = np.interp(action_t, t_idxs, yaws)
   psi_rate = yaw_rates[0]
   return curv_from_psis(psi_target, psi_rate, vego, action_t)

selfdrive/controls/lib/lateral_mpc_lib/SConscript

@@ -1,4 +1,4 @@
-Import('env', 'envCython', 'arch', 'msgq_python', 'common_python', 'np_version')
+Import('env', 'envCython', 'arch', 'msgq_python', 'common_python', 'np_version', 'acados')
 
 gen = "c_generated_code"
 
@@ -45,18 +45,24 @@ generated_files = [
   f'{gen}/lat_cost/lat_cost.h',
 ] + build_files
 
-acados_dir = '#third_party/acados'
-acados_templates_dir = '#third_party/acados/acados_template/c_templates_tera'
+acados_include_dir = Dir(acados.INCLUDE_DIR)
+acados_template_dir = Dir(acados.TEMPLATE_DIR)
 
 source_list = ['lat_mpc.py',
   '#selfdrive/modeld/constants.py',
-  f'{acados_dir}/include/acados_c/ocp_nlp_interface.h',
-  f'{acados_templates_dir}/acados_solver.in.c',
+  acados_include_dir.File('acados_c/ocp_nlp_interface.h'),
+  acados_template_dir.File('c_templates_tera/acados_solver.in.c'),
 ]
 
 lenv = env.Clone()
-acados_rel_path = Dir(gen).rel_path(Dir(f"#third_party/acados/{arch}/lib"))
-lenv["RPATH"] += [lenv.Literal(f'\\$$ORIGIN/{acados_rel_path}')]
+copied_acados_libs = []
+if arch != "Darwin":
+  for lib in ["libacados.so", "libblasfeo.so", "libhpipm.so", "libqpOASES_e.so.3.1"]:
+    copied_acados_libs += lenv.Command(f"{gen}/{lib}", Dir(acados.LIB_DIR).File(lib), [Mkdir(Dir(gen)), Copy("$TARGET", "$SOURCE")])
+  lenv["RPATH"] += [lenv.Literal('\\$$ORIGIN')]
+else:
+  acados_rel_path = Dir(gen).rel_path(Dir(acados.LIB_DIR))
+  lenv["RPATH"] += [lenv.Literal(f'\\$$ORIGIN/{acados_rel_path}')]
 lenv.Clean(generated_files, Dir(gen))
 
 generated_lat = lenv.Command(generated_files,
@@ -77,8 +83,8 @@ lib_solver = lenv.SharedLibrary(f"{gen}/acados_ocp_solver_lat",
                                 LIBS=['m', 'acados', 'hpipm', 'blasfeo', 'qpOASES_e'])
 
 # generate cython stuff
-acados_ocp_solver_pyx = File("#third_party/acados/acados_template/acados_ocp_solver_pyx.pyx")
-acados_ocp_solver_common = File("#third_party/acados/acados_template/acados_solver_common.pxd")
+acados_ocp_solver_pyx = acados_template_dir.File('acados_ocp_solver_pyx.pyx')
+acados_ocp_solver_common = acados_template_dir.File('acados_solver_common.pxd')
 libacados_ocp_solver_pxd = File(f'{gen}/acados_solver.pxd')
 libacados_ocp_solver_c = File(f'{gen}/acados_ocp_solver_pyx.c')
 
@@ -94,4 +100,5 @@ lenv2.Command(libacados_ocp_solver_c,
   f' {acados_ocp_solver_pyx.get_labspath()}')
 lib_cython = lenv2.Program(f'{gen}/acados_ocp_solver_pyx.so', [libacados_ocp_solver_c], LIBS=['acados_ocp_solver_lat'])
 lenv2.Depends(lib_cython, lib_solver)
+lenv2.Depends(lib_cython, copied_acados_libs)
 lenv2.Depends(libacados_ocp_solver_c, np_version)

selfdrive/controls/lib/lateral_mpc_lib/lat_mpc.py

@@ -8,7 +8,7 @@ from casadi import SX, vertcat, sin, cos
 from openpilot.selfdrive.modeld.constants import ModelConstants
 
 if __name__ == '__main__':  # generating code
-  from openpilot.third_party.acados.acados_template import AcadosModel, AcadosOcp, AcadosOcpSolver
+  from acados.acados_template import AcadosModel, AcadosOcp, AcadosOcpSolver
 else:
   from openpilot.selfdrive.controls.lib.lateral_mpc_lib.c_generated_code.acados_ocp_solver_pyx import AcadosOcpSolverCython
 

selfdrive/controls/lib/longcontrol.py

@@ -4,6 +4,7 @@ from openpilot.common.realtime import DT_CTRL
 from openpilot.selfdrive.controls.lib.drive_helpers import CONTROL_N
 from openpilot.common.pid import PIDController
 from openpilot.selfdrive.modeld.constants import ModelConstants
+from openpilot.sunnypilot.selfdrive.controls.lib.brake_smoother import BrakeOnsetSmoother
 
 CONTROL_N_T_IDX = ModelConstants.T_IDXS[:CONTROL_N]
 
@@ -56,6 +57,7 @@ class LongControl:
                              (CP.longitudinalTuning.kiBP, CP.longitudinalTuning.kiV),
                              rate=1 / DT_CTRL)
     self.last_output_accel = 0.0
+    self.brake_smoother = BrakeOnsetSmoother()
 
   def reset(self):
     self.pid.reset()
@@ -87,6 +89,8 @@ class LongControl:
       error = a_target - CS.aEgo
       output_accel = self.pid.update(error, speed=CS.vEgo,
                                      feedforward=a_target)
+      # Ease the brake-onset command (comfort); bypassed for hard braking. See BrakeOnsetSmoother.
+      output_accel = self.brake_smoother.apply(output_accel, self.last_output_accel, a_target)
 
     self.last_output_accel = np.clip(output_accel, accel_limits[0], accel_limits[1])
     return self.last_output_accel

selfdrive/controls/lib/longitudinal_mpc_lib/SConscript

@@ -1,4 +1,4 @@
-Import('env', 'envCython', 'arch', 'msgq_python', 'common_python', 'np_version')
+Import('env', 'envCython', 'arch', 'msgq_python', 'common_python', 'np_version', 'acados')
 
 gen = "c_generated_code"
 
@@ -51,18 +51,24 @@ generated_files = [
   f'{gen}/long_cost/long_cost.h',
 ] + build_files
 
-acados_dir = '#third_party/acados'
-acados_templates_dir = '#third_party/acados/acados_template/c_templates_tera'
+acados_include_dir = Dir(acados.INCLUDE_DIR)
+acados_template_dir = Dir(acados.TEMPLATE_DIR)
 
 source_list = ['long_mpc.py',
   '#selfdrive/modeld/constants.py',
-  f'{acados_dir}/include/acados_c/ocp_nlp_interface.h',
-  f'{acados_templates_dir}/acados_solver.in.c',
+  acados_include_dir.File('acados_c/ocp_nlp_interface.h'),
+  acados_template_dir.File('c_templates_tera/acados_solver.in.c'),
 ]
 
 lenv = env.Clone()
-acados_rel_path = Dir(gen).rel_path(Dir(f"#third_party/acados/{arch}/lib"))
-lenv["RPATH"] += [lenv.Literal(f'\\$$ORIGIN/{acados_rel_path}')]
+copied_acados_libs = []
+if arch != "Darwin":
+  for lib in ["libacados.so", "libblasfeo.so", "libhpipm.so", "libqpOASES_e.so.3.1"]:
+    copied_acados_libs += lenv.Command(f"{gen}/{lib}", Dir(acados.LIB_DIR).File(lib), [Mkdir(Dir(gen)), Copy("$TARGET", "$SOURCE")])
+  lenv["RPATH"] += [lenv.Literal('\\$$ORIGIN')]
+else:
+  acados_rel_path = Dir(gen).rel_path(Dir(acados.LIB_DIR))
+  lenv["RPATH"] += [lenv.Literal(f'\\$$ORIGIN/{acados_rel_path}')]
 lenv.Clean(generated_files, Dir(gen))
 generated_long = lenv.Command(generated_files,
                               source_list,
@@ -82,8 +88,8 @@ lib_solver = lenv.SharedLibrary(f"{gen}/acados_ocp_solver_long",
                                 LIBS=['m', 'acados', 'hpipm', 'blasfeo', 'qpOASES_e'])
 
 # generate cython stuff
-acados_ocp_solver_pyx = File("#third_party/acados/acados_template/acados_ocp_solver_pyx.pyx")
-acados_ocp_solver_common = File("#third_party/acados/acados_template/acados_solver_common.pxd")
+acados_ocp_solver_pyx = acados_template_dir.File('acados_ocp_solver_pyx.pyx')
+acados_ocp_solver_common = acados_template_dir.File('acados_solver_common.pxd')
 libacados_ocp_solver_pxd = File(f'{gen}/acados_solver.pxd')
 libacados_ocp_solver_c = File(f'{gen}/acados_ocp_solver_pyx.c')
 
@@ -99,4 +105,5 @@ lenv2.Command(libacados_ocp_solver_c,
   f' {acados_ocp_solver_pyx.get_labspath()}')
 lib_cython = lenv2.Program(f'{gen}/acados_ocp_solver_pyx.so', [libacados_ocp_solver_c], LIBS=['acados_ocp_solver_long'])
 lenv2.Depends(lib_cython, lib_solver)
+lenv2.Depends(lib_cython, copied_acados_libs)
 lenv2.Depends(libacados_ocp_solver_c, np_version)

selfdrive/controls/lib/longitudinal_mpc_lib/long_mpc.py

@@ -11,7 +11,7 @@ from openpilot.selfdrive.modeld.constants import index_function
 from openpilot.selfdrive.controls.radard import _LEAD_ACCEL_TAU
 
 if __name__ == '__main__':  # generating code
-  from openpilot.third_party.acados.acados_template import AcadosModel, AcadosOcp, AcadosOcpSolver
+  from acados.acados_template import AcadosModel, AcadosOcp, AcadosOcpSolver
 else:
   from openpilot.selfdrive.controls.lib.longitudinal_mpc_lib.c_generated_code.acados_ocp_solver_pyx import AcadosOcpSolverCython
 
@@ -70,6 +70,9 @@ def get_jerk_factor(personality=log.LongitudinalPersonality.standard):
     raise NotImplementedError("Longitudinal personality not supported")
 
 
+T_FOLLOW_MIN = 0.9  # safety floor for follow time after any sunnypilot tier offset
+
+
 def get_T_FOLLOW(personality=log.LongitudinalPersonality.standard):
   if personality==log.LongitudinalPersonality.relaxed:
     return 1.75
@@ -267,8 +270,9 @@ class LongitudinalMpc:
     for i in range(N):
       self.solver.cost_set(i, 'Zl', Zl)
 
-  def set_weights(self, prev_accel_constraint=True, personality=log.LongitudinalPersonality.standard):
-    jerk_factor = get_jerk_factor(personality)
+  def set_weights(self, prev_accel_constraint=True, personality=log.LongitudinalPersonality.standard, jerk_factor_scale=1.0):
+    # jerk_factor_scale (sunnypilot Acceleration Personality): >1 smooths decel/accel, <1 crisps it.
+    jerk_factor = get_jerk_factor(personality) * jerk_factor_scale
     a_change_cost = A_CHANGE_COST if prev_accel_constraint else 0
     cost_weights = [X_EGO_OBSTACLE_COST, X_EGO_COST, V_EGO_COST, A_EGO_COST, jerk_factor * a_change_cost, jerk_factor * J_EGO_COST]
     constraint_cost_weights = [LIMIT_COST, LIMIT_COST, LIMIT_COST, DANGER_ZONE_COST]
@@ -313,8 +317,10 @@ class LongitudinalMpc:
     lead_xv = self.extrapolate_lead(x_lead, v_lead, a_lead, a_lead_tau)
     return lead_xv
 
-  def update(self, radarstate, v_cruise, personality=log.LongitudinalPersonality.standard):
-    t_follow = get_T_FOLLOW(personality)
+  def update(self, radarstate, v_cruise, personality=log.LongitudinalPersonality.standard, t_follow_offset=0.0):
+    # t_follow_offset (sunnypilot Acceleration Personality): + = earlier/gentler decel (bigger gap).
+    # Clamped to a safe minimum so a tier offset can never produce an unsafe follow distance.
+    t_follow = max(T_FOLLOW_MIN, get_T_FOLLOW(personality) + t_follow_offset)
     v_ego = self.x0[1]
     self.status = radarstate.leadOne.status or radarstate.leadTwo.status
 

selfdrive/controls/lib/longitudinal_planner.py

@@ -110,7 +110,7 @@ class LongitudinalPlanner(LongitudinalPlannerSP):
     # No change cost when user is controlling the speed, or when standstill
     prev_accel_constraint = not (reset_state or sm['carState'].standstill)
 
-    accel_clip = [ACCEL_MIN, get_max_accel(v_ego)]
+    accel_clip = [ACCEL_MIN, self.accel.get_max_accel(v_ego)]
     steer_angle_without_offset = sm['carState'].steeringAngleDeg - sm['liveParameters'].angleOffsetDeg
     accel_clip = limit_accel_in_turns(v_ego, steer_angle_without_offset, accel_clip, self.CP)
 
@@ -136,9 +136,12 @@ class LongitudinalPlanner(LongitudinalPlannerSP):
     if force_slow_decel:
       v_cruise = 0.0
 
-    self.mpc.set_weights(prev_accel_constraint, personality=sm['selfdriveState'].personality)
+    # Acceleration Personality decel shaping (ACC path only; stock when disabled or e2e/blended).
+    jerk_factor_scale, t_follow_offset = self.accel.decel_mod(self.is_e2e(sm))
+    self.mpc.set_weights(prev_accel_constraint, personality=sm['selfdriveState'].personality, jerk_factor_scale=jerk_factor_scale)
     self.mpc.set_cur_state(self.v_desired_filter.x, self.a_desired)
-    self.mpc.update(sm['radarState'], v_cruise, personality=sm['selfdriveState'].personality)
+    radarstate = self.smooth_radarstate(sm['radarState'])
+    self.mpc.update(radarstate, v_cruise, personality=sm['selfdriveState'].personality, t_follow_offset=t_follow_offset)
 
     self.v_desired_trajectory = np.interp(CONTROL_N_T_IDX, T_IDXS_MPC, self.mpc.v_solution)
     self.a_desired_trajectory = np.interp(CONTROL_N_T_IDX, T_IDXS_MPC, self.mpc.a_solution)
@@ -169,8 +172,10 @@ class LongitudinalPlanner(LongitudinalPlannerSP):
       output_a_target = output_a_target_mpc
       self.output_should_stop = output_should_stop_mpc
 
-    for idx in range(2):
-      accel_clip[idx] = np.clip(accel_clip[idx], self.prev_accel_clip[idx] - 0.05, self.prev_accel_clip[idx] + 0.05)
+    # Lower (braking) bound and the ceiling's downward slew stay at the stock rate; only the ceiling's
+    # upward slew is tier-dependent (Acceleration Personality).
+    accel_clip[0] = np.clip(accel_clip[0], self.prev_accel_clip[0] - 0.05, self.prev_accel_clip[0] + 0.05)
+    accel_clip[1] = np.clip(accel_clip[1], self.prev_accel_clip[1] - 0.05, self.prev_accel_clip[1] + self.accel.get_rise_rate())
     self.output_a_target = np.clip(output_a_target, accel_clip[0], accel_clip[1])
     self.prev_accel_clip = accel_clip
 

selfdrive/controls/radard.py

@@ -142,7 +142,7 @@ def match_vision_to_track(v_ego: float, lead: capnp._DynamicStructReader, tracks
     return None
 
 
-def get_RadarState_from_vision(lead_msg: capnp._DynamicStructReader, v_ego: float, model_v_ego: float):
+def get_RadarState_from_vision(lead_msg: capnp._DynamicStructReader, v_ego: float, model_v_ego: float, lead_prob: float):
   lead_v_rel_pred = lead_msg.v[0] - model_v_ego
   return {
     "dRel": float(lead_msg.x[0] - RADAR_TO_CAMERA),
@@ -153,7 +153,7 @@ def get_RadarState_from_vision(lead_msg: capnp._DynamicStructReader, v_ego: floa
     "aLeadK": float(lead_msg.a[0]),
     "aLeadTau": 0.3,
     "fcw": False,
-    "modelProb": float(lead_msg.prob),
+    "modelProb": float(lead_prob),
     "status": True,
     "radar": False,
     "radarTrackId": -1,
@@ -161,19 +161,20 @@ def get_RadarState_from_vision(lead_msg: capnp._DynamicStructReader, v_ego: floa
 
 
 def get_lead(v_ego: float, ready: bool, tracks: dict[int, Track], lead_msg: capnp._DynamicStructReader,
-             model_v_ego: float, CP: structs.CarParams, CP_SP: structs.CarParamsSP, low_speed_override: bool = True) -> dict[str, Any]:
+             model_v_ego: float, lead_prob: float, CP: structs.CarParams, CP_SP: structs.CarParamsSP,
+             low_speed_override: bool = True) -> dict[str, Any]:
   # Determine leads, this is where the essential logic happens
-  if len(tracks) > 0 and ready and lead_msg.prob > .5:
+  if len(tracks) > 0 and ready and lead_prob > .5:
     track = match_vision_to_track(v_ego, lead_msg, tracks)
   else:
     track = None
 
   lead_dict = {'status': False}
   if track is not None:
-    lead_dict = track.get_RadarState(lead_msg.prob)
+    lead_dict = track.get_RadarState(lead_prob)
     lead_dict = get_custom_yrel(CP, CP_SP, lead_dict, lead_msg)
-  elif (track is None) and ready and (lead_msg.prob > .5):
-    lead_dict = get_RadarState_from_vision(lead_msg, v_ego, model_v_ego)
+  elif (track is None) and ready and (lead_prob > .5):
+    lead_dict = get_RadarState_from_vision(lead_msg, v_ego, model_v_ego, lead_prob)
 
   if low_speed_override:
     low_speed_tracks = [c for c in tracks.values() if c.potential_low_speed_lead(v_ego)]
@@ -205,6 +206,7 @@ class RadarD:
 
     self.tracks: dict[int, Track] = {}
     self.kalman_params = KalmanParams(DT_MDL)
+    self.lead_prob_filters = [FirstOrderFilter(0.0, 0.2, DT_MDL) for _ in range(2)]
 
     self.v_ego = 0.0
     self.v_ego_hist = deque([0.0], maxlen=int(round(delay / DT_MDL))+1)
@@ -256,8 +258,18 @@ class RadarD:
       model_v_ego = self.v_ego
     leads_v3 = sm['modelV2'].leadsV3
     if len(leads_v3) > 1:
-      self.radar_state.leadOne = get_lead(self.v_ego, self.ready, self.tracks, leads_v3[0], model_v_ego, self.CP, self.CP_SP, low_speed_override=True)
-      self.radar_state.leadTwo = get_lead(self.v_ego, self.ready, self.tracks, leads_v3[1], model_v_ego, self.CP, self.CP_SP, low_speed_override=False)
+      for i in range(2):
+        # Asymmetric filter on lead prob to keep lead when uncertain
+        lead_prob = leads_v3[i].prob
+        if lead_prob > self.lead_prob_filters[i].x:
+          self.lead_prob_filters[i].x = lead_prob
+        else:
+          self.lead_prob_filters[i].update(lead_prob)
+
+      self.radar_state.leadOne = get_lead(self.v_ego, self.ready, self.tracks, leads_v3[0], model_v_ego, self.lead_prob_filters[0].x,
+                                          self.CP, self.CP_SP, low_speed_override=True)
+      self.radar_state.leadTwo = get_lead(self.v_ego, self.ready, self.tracks, leads_v3[1], model_v_ego, self.lead_prob_filters[1].x,
+                                          self.CP, self.CP_SP, low_speed_override=False)
 
   def publish(self, pm: messaging.PubMaster):
     assert self.radar_state is not None

selfdrive/debug/car/ecu_addrs.py

@@ -26,9 +26,9 @@ if __name__ == "__main__":
   # Set up params for pandad
   params = Params()
   params.remove("FirmwareQueryDone")
-  params.put_bool("IsOnroad", False)
+  params.put_bool("IsOnroad", False, block=True)
   time.sleep(0.2)  # thread is 10 Hz
-  params.put_bool("IsOnroad", True)
+  params.put_bool("IsOnroad", True, block=True)
 
   obd_callback(params)(not args.no_obd)
 

selfdrive/debug/car/fw_versions.py

@@ -30,9 +30,9 @@ if __name__ == "__main__":
   # Set up params for pandad
   params = Params()
   params.remove("FirmwareQueryDone")
-  params.put_bool("IsOnroad", False)
+  params.put_bool("IsOnroad", False, block=True)
   time.sleep(0.2)  # thread is 10 Hz
-  params.put_bool("IsOnroad", True)
+  params.put_bool("IsOnroad", True, block=True)
   set_obd_multiplexing = obd_callback(params)
 
   extra: Any = None

selfdrive/debug/set_car_params.py

@@ -19,4 +19,4 @@ if __name__ == "__main__":
 
   cp_bytes = CP.to_bytes()
   for p in ("CarParams", "CarParamsCache", "CarParamsPersistent"):
-    Params().put(p, cp_bytes)
+    Params().put(p, cp_bytes, block=True)

selfdrive/debug/uiview.py

@@ -9,7 +9,7 @@ from openpilot.system.hardware import HARDWARE
 if __name__ == "__main__":
   CP = car.CarParams(notCar=True, wheelbase=1, steerRatio=10)
   params = Params()
-  params.put("CarParams", CP.to_bytes())
+  params.put("CarParams", CP.to_bytes(), block=True)
 
   if use_tinygrad_modeld := is_tinygrad_model(False, params, CP):
     print("Using TinyGrad modeld")

selfdrive/locationd/calibrationd.py

@@ -167,7 +167,7 @@ class Calibrator:
 
     write_this_cycle = (self.idx == 0) and (self.block_idx % (INPUTS_WANTED//5) == 5)
     if self.param_put and write_this_cycle:
-      self.params.put_nonblocking("CalibrationParams", self.get_msg(True).to_bytes())
+      self.params.put("CalibrationParams", self.get_msg(True).to_bytes())
 
   def handle_v_ego(self, v_ego: float) -> None:
     self.v_ego = v_ego

selfdrive/locationd/lagd.py

@@ -414,7 +414,7 @@ def main():
       pm.send('liveDelay', lag_msg_dat)
 
       if sm.frame % 1200 == 0: # cache every 60 seconds
-        params.put_nonblocking("LiveDelay", lag_msg_dat)
+        params.put("LiveDelay", lag_msg_dat)
 
       if sm.frame % 60 == 0:  # read from and write to params every 3 seconds
         lagd_toggle.update(lag_msg)

selfdrive/locationd/paramsd.py

@@ -212,7 +212,7 @@ def migrate_cached_vehicle_params_if_needed(params: Params):
     last_parameters_msg.liveParameters.steerRatio = last_parameters_data_old['steerRatio']
     last_parameters_msg.liveParameters.stiffnessFactor = last_parameters_data_old['stiffnessFactor']
     last_parameters_msg.liveParameters.angleOffsetAverageDeg = last_parameters_data_old['angleOffsetAverageDeg']
-    params.put("LiveParametersV2", last_parameters_msg.to_bytes())
+    params.put("LiveParametersV2", last_parameters_msg.to_bytes(), block=True)
   except Exception as e:
     cloudlog.error(f"Failed to perform parameter migration: {e}")
     params.remove("LiveParameters")
@@ -290,7 +290,7 @@ def main():
 
       msg_dat = msg.to_bytes()
       if sm.frame % 1200 == 0:  # once a minute
-        params.put_nonblocking("LiveParametersV2", msg_dat)
+        params.put("LiveParametersV2", msg_dat)
 
       pm.send('liveParameters', msg_dat)
 

selfdrive/locationd/test/test_calibrationd.py

@@ -36,7 +36,7 @@ class TestCalibrationd:
     msg.liveCalibration.validBlocks = random.randint(1, 10)
     msg.liveCalibration.rpyCalib = [random.random() for _ in range(3)]
     msg.liveCalibration.height = [random.random() for _ in range(1)]
-    Params().put("CalibrationParams", msg.to_bytes())
+    Params().put("CalibrationParams", msg.to_bytes(), block=True)
     c = Calibrator(param_put=True)
 
     np.testing.assert_allclose(msg.liveCalibration.rpyCalib, c.rpy)

selfdrive/locationd/test/test_lagd.py

@@ -53,8 +53,8 @@ class TestLagd:
     msg = messaging.new_message('liveDelay')
     msg.liveDelay.lateralDelayEstimate = random.random()
     msg.liveDelay.validBlocks = random.randint(1, 10)
-    params.put("LiveDelay", msg.to_bytes())
-    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes())
+    params.put("LiveDelay", msg.to_bytes(), block=True)
+    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes(), block=True)
 
     saved_lag_params = retrieve_initial_lag(params, CP)
     assert saved_lag_params is not None

selfdrive/locationd/test/test_paramsd.py

@@ -27,8 +27,8 @@ class TestParamsd:
     CP = next(m for m in lr if m.which() == "carParams").carParams
 
     msg = get_random_live_parameters(CP)
-    params.put("LiveParametersV2", msg.to_bytes())
-    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes())
+    params.put("LiveParametersV2", msg.to_bytes(), block=True)
+    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes(), block=True)
 
     migrate_cached_vehicle_params_if_needed(params) # this is not tested here but should not mess anything up or throw an error
     sr, sf, offset, p_init = retrieve_initial_vehicle_params(params, CP, replay=True, debug=True)
@@ -46,8 +46,8 @@ class TestParamsd:
     CP = next(m for m in lr if m.which() == "carParams").carParams
 
     msg = get_random_live_parameters(CP)
-    params.put("LiveParameters", msg.liveParameters.to_dict())
-    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes())
+    params.put("LiveParameters", msg.liveParameters.to_dict(), block=True)
+    params.put("CarParamsPrevRoute", CP.as_builder().to_bytes(), block=True)
     params.remove("LiveParametersV2")
 
     migrate_cached_vehicle_params_if_needed(params)
@@ -59,7 +59,7 @@ class TestParamsd:
 
   def test_read_saved_params_corrupted_old_format(self):
     params = Params()
-    params.put("LiveParameters", {})
+    params.put("LiveParameters", {}, block=True)
     params.remove("LiveParametersV2")
 
     migrate_cached_vehicle_params_if_needed(params)

selfdrive/locationd/torqued.py

@@ -278,7 +278,7 @@ def main(demo=False):
     # Cache points every 60 seconds while onroad
     if sm.frame % 240 == 0:
       msg = estimator.get_msg(valid=sm.all_checks(), with_points=True)
-      params.put_nonblocking("LiveTorqueParameters", msg.to_bytes())
+      params.put("LiveTorqueParameters", msg.to_bytes())
 
 
 if __name__ == "__main__":

selfdrive/modeld/SConscript

@@ -1,6 +1,19 @@
-import os
 import glob
-from openpilot.common.file_chunker import chunk_file, get_chunk_paths
+import json
+import os
+import sys, subprocess
+from SCons.Script import Action, Value
+from openpilot.common.file_chunker import chunk_file, get_chunk_targets, get_existing_chunks
+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 TG_INPUT_DEVICES_PATH, usbgpu_present, modeld_pkl_path
+
+
+CAMERA_CONFIGS = [
+  (_ar_ox_fisheye.width, _ar_ox_fisheye.height),  # tici: 1928x1208
+  (_os_fisheye.width, _os_fisheye.height),        # mici: 1344x760
+]
 
 Import('env', 'arch')
 chunker_file = File("#common/file_chunker.py")
@@ -13,52 +26,116 @@ 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
 
-# compile warp
-# THREADS=0 is need to prevent bug: https://github.com/tinygrad/tinygrad/issues/14689
-tg_flags = {
-    'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0',
-    'Darwin': f'DEV=CPU THREADS=0 HOME={os.path.expanduser("~")}', # tinygrad calls brew which needs a $HOME in the env
-}.get(arch, 'DEV=CPU CPU_LLVM=1 THREADS=0')
+# get fastest TG config
+# probe in subprocess so usbgpu locks gets released on process exit
+def probe_devices():
+  return set(subprocess.run(
+    [sys.executable, '-c', 'from tinygrad import Device\nprint("\\n".join(Device.get_available_devices()))'],
+    capture_output=True, text=True, check=True).stdout.strip().splitlines())
 
-# Get model metadata
-for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
-  fn = File(f"models/{model_name}").abspath
-  script_files = [File(Dir("#selfdrive/modeld").File("get_model_metadata.py").abspath)]
-  cmd = f'{tg_flags} python3 {Dir("#selfdrive/modeld").abspath}/get_model_metadata.py {fn}.onnx'
-  lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files + script_files, cmd)
+available = probe_devices()
+if '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} IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1'
+else:
+  tg_backend = 'CPU'
+  tg_flags = f'DEV=CPU' if arch == 'Darwin' else 'DEV=CPU:LLVM'
 
-image_flag = {
-     'larch64': 'IMAGE=2',
-}.get(arch, 'IMAGE=0')
-script_files = [File(Dir("#selfdrive/modeld").File("compile_warp.py").abspath)]
-compile_warp_cmd = f'{tg_flags} python3 {Dir("#selfdrive/modeld").abspath}/compile_warp.py '
-from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
-warp_targets = []
-for cam in [_ar_ox_fisheye, _os_fisheye]:
-  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]
-lenv.Command(warp_targets, tinygrad_files + script_files, compile_warp_cmd)
+tg_devices = { # which device to put jit inputs to at runtime
+  'selfdrive.modeld.modeld': {
+    'default': {'WARP_DEV': tg_backend, 'QUEUE_DEV': tg_backend},
+    'usbgpu': {'WARP_DEV': tg_backend, 'QUEUE_DEV': 'AMD'}
+  },
+  'selfdrive.modeld.dmonitoringmodeld': {
+    'default': {'DEV': tg_backend}
+  },
+}
+
+USBGPU = usbgpu_present() # or release # TODO always build big model on release
+if USBGPU:
+  usbgpu_tg_flags = f'DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV={tg_backend} FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0'
+  # the USB+AMD GPU takes an exclusive flock; serialize all targets that touch it
+  usbgpu_lock = File("models/.usb_gpu.lock").abspath
+
+def write_tg_devices(target, source, env):
+  with open(str(target[0]), "w") as f:
+    json.dump(tg_devices, f)
+    f.write("\n")
+
+tg_devices_node = lenv.Command(
+  str(TG_INPUT_DEVICES_PATH),
+  [Value(tg_devices)],
+  write_tg_devices,
+)
+
+# tinygrad calls brew which needs a $HOME in the env
+mac_brew_string = f'HOME={os.path.expanduser("~")}' if arch == 'Darwin' else ''
+
+modeld_dir = Dir("#selfdrive/modeld").abspath
+compile_modeld_script = [
+  File(f"{modeld_dir}/compile_modeld.py"),
+  File(f"{modeld_dir}/get_model_metadata.py"),
+  File("#system/camerad/cameras/nv12_info.py"),
+  File("#system/hardware/hw.py"),
+]
+model_w, model_h = MEDMODEL_INPUT_SIZE
+frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
+
+for usbgpu in [False, True] if USBGPU else [False]:
+  target_pkl_path = File(modeld_pkl_path(usbgpu)).abspath
+  file_prefix, cmd_flags = ('big_', usbgpu_tg_flags) if usbgpu else ('', tg_flags)
+  driving_onnx_deps = [p for m in [f'{file_prefix}driving_vision', f'{file_prefix}driving_on_policy']
+                       for p in get_existing_chunks(File(f"models/{m}.onnx").abspath)]
+  camera_res_args = ' '.join(f'{cw}x{ch}' for cw, ch in CAMERA_CONFIGS)
+  cmd = (f'{cmd_flags} {mac_brew_string} python3 {modeld_dir}/compile_modeld.py '
+        f'--model-size {model_w}x{model_h} '
+        f'--camera-resolutions {camera_res_args} '
+        f'--vision-onnx {File(f"models/{file_prefix}driving_vision.onnx").abspath} '
+        f'--on-policy-onnx {File(f"models/{file_prefix}driving_on_policy.onnx").abspath} '
+        f'--output {target_pkl_path} --frame-skip {frame_skip}')
+  onnx_sizes_sum = sum(os.path.getsize(f) for f in driving_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)
+  node = lenv.Command(
+    chunk_targets,
+    tinygrad_files + compile_modeld_script + driving_onnx_deps + [Value(chunk_targets), chunker_file],
+    [cmd, Action(do_chunk, "  [CHUNK] $TARGET")],
+  )
+  if usbgpu:
+    lenv.SideEffect(usbgpu_lock, node)
+
+# get model metadata
+fn = File(f"models/dmonitoring_model").abspath
+script_files = [File(Dir("#selfdrive/modeld").File("get_model_metadata.py").abspath)]
+cmd = f'{tg_flags} {mac_brew_string} python3 {Dir("#selfdrive/modeld").abspath}/get_model_metadata.py {fn}.onnx'
+lenv.Command(fn + "_metadata.pkl", [fn + ".onnx"] + tinygrad_files + script_files + [tg_devices_node], cmd)
+
+dm_w, dm_h = DM_INPUT_SIZE
+compile_dm_warp_script = [File(f"{modeld_dir}/compile_dm_warp.py")]
+for cam_w, cam_h in CAMERA_CONFIGS:
+  dm_pkl_path = File(f"models/dm_warp_{cam_w}x{cam_h}_tinygrad.pkl").abspath
+  cmd = (f'{tg_flags} {mac_brew_string} python3 {modeld_dir}/compile_dm_warp.py '
+         f'--camera-resolution {cam_w}x{cam_h} --warp-to {dm_w}x{dm_h} '
+         f'--output {dm_pkl_path}')
+  lenv.Command(dm_pkl_path, tinygrad_files + compile_dm_warp_script + compile_modeld_script + [tg_devices_node], cmd)
 
 def tg_compile(flags, model_name):
   pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + '"'
   fn = File(f"models/{model_name}").abspath
   pkl = fn + "_tinygrad.pkl"
   onnx_path = fn + ".onnx"
-  chunk_targets = get_chunk_paths(pkl, estimate_pickle_max_size(os.path.getsize(onnx_path)))
-  compile_node = lenv.Command(
-    pkl,
-    [onnx_path] + tinygrad_files + [chunker_file],
-    f'{pythonpath_string} {flags} {image_flag} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {pkl}',
-  )
+  chunk_targets = get_chunk_targets(pkl, estimate_pickle_max_size(os.path.getsize(onnx_path)))
   def do_chunk(target, source, env):
     chunk_file(pkl, chunk_targets)
   return lenv.Command(
     chunk_targets,
-    compile_node,
-    do_chunk,
+    [onnx_path] + tinygrad_files + [Value(chunk_targets), chunker_file, tg_devices_node],
+    [f'{pythonpath_string} {flags} python3 {Dir("#tinygrad_repo").abspath}/examples/openpilot/compile3.py {fn}.onnx {pkl}',
+     Action(do_chunk, "  [CHUNK] $TARGET")],
   )
 
-# Compile small models
-for model_name in ['driving_vision', 'driving_policy', 'dmonitoring_model']:
-  tg_compile(tg_flags, model_name)
-
+tg_compile(tg_flags, 'dmonitoring_model')

selfdrive/modeld/compile_dm_warp.py

@@ -0,0 +1,56 @@
+#!/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.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.selfdrive.modeld.compile_modeld import NV12Frame, warp_perspective_tinygrad, _parse_size
+
+
+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, border_fill_val=16).reshape(-1, dm_h * dm_w) # Y
+  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('--camera-resolution', type=_parse_size, required=True, help='camera resolution WxH')
+  p.add_argument('--warp-to', type=_parse_size, required=True, help='DM input WxH')
+  p.add_argument('--output', required=True)
+  args = p.parse_args()
+
+  cam_w, cam_h = args.camera_resolution
+  nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+  dm_w, dm_h = args.warp_to
+  compile_dm_warp(nv12, dm_w, dm_h, args.output)

selfdrive/modeld/compile_modeld.py

@@ -0,0 +1,299 @@
+#!/usr/bin/env python3
+import argparse
+import atexit
+import os
+import pickle
+import time
+from functools import partial
+from collections import namedtuple, defaultdict
+
+import numpy as np
+
+def _patch_tinygrad_fetch_fw():
+  import hashlib
+  import pathlib
+  import zstandard
+  from tinygrad import helpers
+  _orig = helpers.fetch_fw
+  def fetch_fw(path, name, sha256):
+    p = pathlib.Path(f"/lib/firmware/{path}/{name}.zst")
+    if p.is_file():
+      blob = zstandard.ZstdDecompressor().stream_reader(p.read_bytes()).read()
+      if hashlib.sha256(blob).hexdigest() == sha256:
+        return blob
+    return _orig(path, name, sha256)
+  helpers.fetch_fw = fetch_fw
+_patch_tinygrad_fetch_fw()
+
+from tinygrad.tensor import Tensor
+from tinygrad.helpers import Context
+from tinygrad.device import Device
+from tinygrad.engine.jit import TinyJit
+
+
+NV12Frame = namedtuple("NV12Frame", ['width', 'height', 'stride', 'y_height', 'uv_height', 'size'])
+WARP_INPUTS = ['img_q', 'big_img_q', 'tfm', 'big_tfm']
+POLICY_INPUTS = ['feat_q', 'desire_q', 'desire', 'traffic_convention', 'action_t']
+
+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)
+
+WARP_DEV = os.getenv('WARP_DEV')
+
+
+def make_random_images(keys, shape, device=None):
+  return {k: Tensor.randint(shape, low=0, high=256, dtype='uint8', device=device).realize() for k in keys}
+
+
+def warp_perspective_tinygrad(src_flat, M_inv, dst_shape, src_shape, stride_pad, border_fill_val=None):
+  w_dst, h_dst = dst_shape
+  h_src, w_src = src_shape
+
+  x = Tensor.arange(w_dst, device=WARP_DEV).reshape(1, w_dst).expand(h_dst, w_dst).reshape(-1)
+  y = Tensor.arange(h_dst, device=WARP_DEV).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_round = Tensor.round(src_x)
+  y_round = Tensor.round(src_y)
+  x_nn_clipped = x_round.clip(0, w_src - 1).cast('int')
+  y_nn_clipped = y_round.clip(0, h_src - 1).cast('int')
+  idx = y_nn_clipped * (w_src + stride_pad) + x_nn_clipped
+  sampled = src_flat[idx]
+
+  if border_fill_val is None:
+    return sampled
+
+  in_bounds = ((x_round >= 0) & (x_round <= w_src - 1) &
+               (y_round >= 0) & (y_round <= h_src - 1)).cast(sampled.dtype)
+  return sampled * in_bounds + Tensor(border_fill_val, dtype=sampled.dtype) * (1 - in_bounds)
+
+
+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]], device=WARP_DEV)
+    # 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, device):
+  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)
+  #TODO action_t is hardcoded to match tc for future compatibility
+  at = tc
+
+  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),
+    'action_t': np.zeros(at, dtype=np.float32),
+  }
+  input_queues = {
+    'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    'feat_q': Tensor(np.zeros((frame_skip * (fb[1] - 1) + 1, fb[0], fb[2]), dtype=np.float32), device=device).contiguous().realize(),
+    'desire_q': Tensor(np.zeros((frame_skip * dp[1], dp[0], dp[2]), dtype=np.float32), device=device).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_warp(nv12, model_w, model_h, frame_skip):
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
+  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
+
+  def warp_enqueue(img_q, big_img_q, tfm, big_tfm, frame, big_frame):
+    tfm = tfm.to(WARP_DEV)
+    big_tfm = big_tfm.to(WARP_DEV)
+    Tensor.realize(tfm, big_tfm)
+
+    warped_frame = frame_prepare(frame, tfm).unsqueeze(0).to(Device.DEFAULT)
+    warped_big_frame = frame_prepare(big_frame, big_tfm).unsqueeze(0).to(Device.DEFAULT)
+    img = shift_and_sample(img_q, warped_frame, sample_skip_fn)
+    big_img = shift_and_sample(big_img_q, warped_big_frame, sample_skip_fn)
+    return img, big_img
+  return warp_enqueue
+
+
+def make_run_policy(vision_runner, on_policy_runner, vision_features_slice, frame_skip):
+  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
+  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
+
+  def run_policy(img, big_img, feat_q, desire_q, desire, traffic_convention, action_t):
+    desire = desire.to(Device.DEFAULT)
+    traffic_convention = traffic_convention.to(Device.DEFAULT)
+    action_t = action_t.to(Device.DEFAULT)
+    Tensor.realize(desire, traffic_convention, action_t)
+    desire_buf = shift_and_sample(desire_q, desire.reshape(1, 1, -1), sample_desire_fn)
+    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)
+
+    inputs = {
+      'features_buffer': feat_buf,
+      'desire_pulse': desire_buf,
+      'traffic_convention': traffic_convention,
+      'action_t': action_t,
+    }
+    on_policy_out = next(iter(on_policy_runner(inputs).values())).cast('float32')
+    #off_policy_out = next(iter(off_policy_runner(inputs).values())).cast('float32')
+    return vision_out, on_policy_out
+
+  return run_policy
+
+
+def compile_jit(jit, make_random_inputs, input_keys, frame_skip, vision_metadata, policy_metadata):
+  vision_input_shapes = vision_metadata['input_shapes']
+  policy_input_shapes = policy_metadata['input_shapes']
+
+  SEED = 42
+  def random_inputs_run(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, Device.DEFAULT)
+    np.random.seed(seed)
+    Tensor.manual_seed(seed)
+
+    testing = test_val is not None or test_buffers is not None
+    n_runs = 1 if testing else 3
+
+    for i in range(n_runs):
+      for v in npy.values():
+        v[:] = np.random.randn(*v.shape).astype(v.dtype)
+      Device.default.synchronize()
+      random_inputs = make_random_inputs()
+      st = time.perf_counter()
+      outs = fn(**{k: input_queues[k] for k in input_keys}, **random_inputs)
+      mt = time.perf_counter()
+      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")
+
+      if i == 0:
+        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 val, buffers
+
+  print('capture + replay')
+  test_val, test_buffers = random_inputs_run(jit, SEED)
+  print('pickle round trip')
+  jit = pickle.loads(pickle.dumps(jit))
+  random_inputs_run(jit, SEED, test_val, test_buffers, expect_match=True)
+  random_inputs_run(jit, SEED+1, test_val, test_buffers, expect_match=False)
+  return jit
+
+
+def _parse_size(s):
+  w, h = s.lower().split('x')
+  return int(w), int(h)
+
+
+def read_file_chunked_to_shm(path):
+  from openpilot.common.file_chunker import read_file_chunked
+  from openpilot.system.hardware.hw import Paths
+  shm_path = os.path.join(Paths.shm_path(), os.path.basename(path))
+  atexit.register(lambda: os.path.exists(shm_path) and os.remove(shm_path))
+  with open(shm_path, 'wb') as f:
+    f.write(read_file_chunked(path))
+  return shm_path
+
+
+if __name__ == "__main__":
+  from tinygrad.nn.onnx import OnnxRunner
+  from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+  from openpilot.selfdrive.modeld.get_model_metadata import make_metadata_dict
+  p = argparse.ArgumentParser()
+  p.add_argument('--model-size', type=_parse_size, required=True, help='model input WxH')
+  p.add_argument('--camera-resolutions', type=_parse_size, nargs='+', required=True,
+                 help='camera resolutions WxH (one or more)')
+  p.add_argument('--vision-onnx', required=True)
+  p.add_argument('--on-policy-onnx', required=True)
+  p.add_argument('--output', required=True)
+  p.add_argument('--frame-skip', type=int, required=True)
+  args = p.parse_args()
+
+  out = defaultdict(dict)
+  vision_path, on_policy_path = read_file_chunked_to_shm(args.vision_onnx), read_file_chunked_to_shm(args.on_policy_onnx)
+  model_w, model_h = args.model_size
+
+  vision_runner = OnnxRunner(vision_path)
+  on_policy_runner = OnnxRunner(on_policy_path)
+  vision_metadata, on_policy_metadata = make_metadata_dict(vision_path), make_metadata_dict(on_policy_path)
+
+  run_policy_jit = TinyJit(make_run_policy(vision_runner, on_policy_runner, vision_metadata['output_slices']['hidden_state'], args.frame_skip), prune=True)
+  out['metadata']['vision'], out['metadata']['on_policy'] = vision_metadata, on_policy_metadata
+
+  make_random_model_inputs = partial(make_random_images, keys=['img', 'big_img'], shape=vision_metadata['input_shapes']['img'])
+  out['run_policy'] = compile_jit(run_policy_jit, make_random_model_inputs, POLICY_INPUTS, args.frame_skip, vision_metadata, on_policy_metadata)
+
+  for cam_w, cam_h in args.camera_resolutions:
+    nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+    make_random_warp_inputs = partial(make_random_images, keys=['frame', 'big_frame'], shape=nv12.size, device=WARP_DEV)
+    warp_enqueue = TinyJit(make_warp(nv12, model_w, model_h, args.frame_skip), prune=True)
+    out[(cam_w,cam_h)] = compile_jit(warp_enqueue, make_random_warp_inputs, WARP_INPUTS, args.frame_skip, vision_metadata, on_policy_metadata)
+
+  with open(args.output, "wb") as f:
+    pickle.dump(out, f)
+  print(f"Saved JITs to {args.output} ({os.path.getsize(args.output) / 1e6:.2f} MB)")

selfdrive/modeld/compile_warp.py

@@ -97,8 +97,8 @@ 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)
-    full_buffer = tensor[6:].cat(new_img, dim=0).contiguous()
-    return full_buffer, Tensor.cat(full_buffer[:6], full_buffer[-6:], dim=0).contiguous().reshape(1, 12, model_h//2, model_w//2)
+    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
 
 
@@ -107,9 +107,9 @@ def make_update_both_imgs(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_buffer, calib_img_pair = update_img(calib_img_buffer, new_img, M_inv)
-    calib_big_img_buffer, calib_big_img_pair = update_img(calib_big_img_buffer, new_big_img, M_inv_big)
-    return calib_img_buffer, calib_img_pair, calib_big_img_buffer, calib_big_img_pair
+    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
 
 
@@ -136,29 +136,20 @@ def compile_modeld_warp(cam_w, cam_h):
 
   full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
   big_full_buffer = Tensor.zeros(IMG_BUFFER_SHAPE, dtype='uint8').contiguous().realize()
-  full_buffer_np = np.zeros(IMG_BUFFER_SHAPE, dtype=np.uint8)
-  big_full_buffer_np = np.zeros(IMG_BUFFER_SHAPE, dtype=np.uint8)
-
+  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):
-    new_frame_np = (32 * np.random.randn(yuv_size).astype(np.float32) + 128).clip(0, 255).astype(np.uint8)
     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')]
-    new_big_frame_np = (32 * np.random.randn(yuv_size).astype(np.float32) + 128).clip(0, 255).astype(np.uint8)
     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()
 
-    inputs_np = [x.numpy() for x in inputs]
-    inputs_np[0] = full_buffer_np
-    inputs_np[3] = big_full_buffer_np
-
     st = time.perf_counter()
-    out = update_img_jit(*inputs)
-    full_buffer = out[0].contiguous().realize().clone()
-    big_full_buffer = out[2].contiguous().realize().clone()
+    _ = update_img_jit(*inputs)
     mt = time.perf_counter()
     Device.default.synchronize()
     et = time.perf_counter()
@@ -182,8 +173,9 @@ def compile_dm_warp(cam_w, 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((32 * Tensor.randn(yuv_size,) + 128).cast(dtype='uint8').realize().numpy().ctypes.data, (yuv_size,), dtype='uint8'),
+    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()

selfdrive/modeld/constants.py

@@ -38,6 +38,7 @@ class ModelConstants:
   LANE_LINES_WIDTH = 2
   ROAD_EDGES_WIDTH = 2
   PLAN_WIDTH = 15
+  ACTION_WIDTH = 2
   DESIRE_PRED_WIDTH = 8
   LAT_PLANNER_SOLUTION_WIDTH = 4
   DESIRED_CURV_WIDTH = 1

selfdrive/modeld/dmonitoringmodeld.py

@@ -1,12 +1,10 @@
 #!/usr/bin/env python3
 import os
-from openpilot.system.hardware import TICI
-os.environ['DEV'] = 'QCOM' if TICI else 'CPU'
+from openpilot.selfdrive.modeld.helpers import MODELS_DIR, get_tg_input_devices
 from tinygrad.tensor import Tensor
 import time
 import pickle
 import numpy as np
-from pathlib import Path
 
 from cereal import messaging
 from cereal.messaging import PubMaster, SubMaster
@@ -21,15 +19,16 @@ from openpilot.selfdrive.modeld.parse_model_outputs import sigmoid, safe_exp
 
 PROCESS_NAME = "selfdrive.modeld.dmonitoringmodeld"
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
-MODEL_PKL_PATH = Path(__file__).parent / 'models/dmonitoring_model_tinygrad.pkl'
-METADATA_PATH = Path(__file__).parent / 'models/dmonitoring_model_metadata.pkl'
-MODELS_DIR = Path(__file__).parent / 'models'
+MODEL_PKL_PATH = MODELS_DIR / 'dmonitoring_model_tinygrad.pkl'
+METADATA_PATH = MODELS_DIR / 'dmonitoring_model_metadata.pkl'
+
 
 class ModelState:
   inputs: dict[str, np.ndarray]
   output: np.ndarray
 
-  def __init__(self):
+  def __init__(self, cam_w: int, cam_h: int):
+    self.DEV = get_tg_input_devices(PROCESS_NAME, usbgpu=False)['DEV']
     with open(METADATA_PATH, 'rb') as f:
       model_metadata = pickle.load(f)
       self.input_shapes = model_metadata['input_shapes']
@@ -41,31 +40,27 @@ 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(MODELS_DIR / f'dm_warp_{cam_w}x{cam_h}_tinygrad.pkl', "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
+    ptr = np.frombuffer(buf.data, dtype=np.uint8).ctypes.data
     # There is a ringbuffer of imgs, just cache tensors pointing to all of them
     if ptr not in self._blob_cache:
-      self._blob_cache[ptr] = Tensor.from_blob(ptr, (self.frame_buf_params[3],), dtype='uint8')
+      self._blob_cache[ptr] = Tensor.from_blob(ptr, (self.frame_buf_params[3],), dtype='uint8', device=self.DEV)
 
     self.warp_inputs_np['transform'][:] = transform[:]
-    self.tensor_inputs['input_img'] = self.image_warp(self._blob_cache[ptr], self.warp_inputs['transform']).realize()
+    self.tensor_inputs['input_img'] = self.image_warp(self._blob_cache[ptr], self.warp_inputs['transform'])
 
-    output = self.model_run(**self.tensor_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
+    output = self.model_run(**self.tensor_inputs).numpy().flatten()
 
     t2 = time.perf_counter()
     return output, t2 - t1
@@ -80,7 +75,7 @@ def parse_model_output(model_output):
     face_descs = model_output[f'face_descs_{ds_suffix}']
     parsed[f'face_descs_{ds_suffix}'] = face_descs[:, :-6]
     parsed[f'face_descs_{ds_suffix}_std'] = safe_exp(face_descs[:, -6:])
-    for key in ['face_prob', 'eyes_visible_prob', 'eyes_closed_prob', 'using_phone_prob']:
+    for key in ['face_prob', 'left_eye_prob', 'right_eye_prob','left_blink_prob', 'right_blink_prob', 'sunglasses_prob', 'using_phone_prob', 'sleep_prob']:
       parsed[f'{key}_{ds_suffix}'] = sigmoid(model_output[f'{key}_{ds_suffix}'])
   return parsed
 
@@ -90,9 +85,13 @@ def fill_driver_data(msg, model_output, ds_suffix):
   msg.facePosition = model_output[f'face_descs_{ds_suffix}'][0, 3:5].tolist()
   msg.facePositionStd = model_output[f'face_descs_{ds_suffix}_std'][0, 3:5].tolist()
   msg.faceProb = model_output[f'face_prob_{ds_suffix}'][0, 0].item()
-  msg.eyesVisibleProb = model_output[f'eyes_visible_prob_{ds_suffix}'][0, 0].item()
-  msg.eyesClosedProb = model_output[f'eyes_closed_prob_{ds_suffix}'][0, 0].item()
+  msg.leftEyeProb = model_output[f'left_eye_prob_{ds_suffix}'][0, 0].item()
+  msg.rightEyeProb = model_output[f'right_eye_prob_{ds_suffix}'][0, 0].item()
+  msg.leftBlinkProb = model_output[f'left_blink_prob_{ds_suffix}'][0, 0].item()
+  msg.rightBlinkProb = model_output[f'right_blink_prob_{ds_suffix}'][0, 0].item()
+  msg.sunglassesProb = model_output[f'sunglasses_prob_{ds_suffix}'][0, 0].item()
   msg.phoneProb = model_output[f'using_phone_prob_{ds_suffix}'][0, 0].item()
+  msg.sleepProb = model_output[f'sleep_prob_{ds_suffix}'][0, 0].item()
 
 def get_driverstate_packet(model_output, frame_id: int, location_ts: int, exec_time: float, gpu_exec_time: float):
   msg = messaging.new_message('driverStateV2', valid=True)
@@ -110,9 +109,6 @@ 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):
@@ -120,6 +116,9 @@ 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

@@ -35,21 +35,21 @@ def get_metadata_value_by_name(model: dict[str, Any], name: str) -> str | Any:
   return None
 
 
-if __name__ == "__main__":
-  model_path = pathlib.Path(sys.argv[1])
+def make_metadata_dict(model_path):
   model = MetadataOnnxPBParser(model_path).parse()
   output_slices = get_metadata_value_by_name(model, 'output_slices')
   assert output_slices is not None, 'output_slices not found in metadata'
-
-  metadata = {
+  return {
     'model_checkpoint': get_metadata_value_by_name(model, 'model_checkpoint'),
     'output_slices': pickle.loads(codecs.decode(output_slices.encode(), "base64")),
     'input_shapes': dict(get_name_and_shape(x) for x in model["graph"]["input"]),
     'output_shapes': dict(get_name_and_shape(x) for x in model["graph"]["output"]),
   }
 
+
+if __name__ == "__main__":
+  model_path = pathlib.Path(sys.argv[1])
   metadata_path = model_path.parent / (model_path.stem + '_metadata.pkl')
   with open(metadata_path, 'wb') as f:
-    pickle.dump(metadata, f)
-
+    pickle.dump(make_metadata_dict(model_path), f)
   print(f'saved metadata to {metadata_path}')

selfdrive/modeld/helpers.py

@@ -0,0 +1,26 @@
+import json
+from pathlib import Path
+
+MODELS_DIR = Path(__file__).resolve().parent / 'models'
+TG_INPUT_DEVICES_PATH = MODELS_DIR / 'tg_input_devices.json'
+USBGPU_VID = 0xADD1
+USBGPU_PID = 0x0001
+
+
+def get_tg_input_devices(process_name: str, usbgpu: bool):
+  with open(TG_INPUT_DEVICES_PATH) as f:
+    return json.load(f)[process_name]['default' if not usbgpu else 'usbgpu']
+
+def modeld_pkl_path(usbgpu: bool):
+  prefix = 'big_' if usbgpu else ''
+  return MODELS_DIR / f'{prefix}driving_tinygrad.pkl'
+
+def usbgpu_present() -> bool:
+  for d in Path("/sys/bus/usb/devices").glob("*"):
+    try:
+      if int((d / "idVendor").read_text(), 16) == USBGPU_VID and \
+          int((d / "idProduct").read_text(), 16) == USBGPU_PID:
+        return True
+    except Exception:
+      pass
+  return False

selfdrive/modeld/modeld.py

@@ -1,18 +1,12 @@
 #!/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'
+os.environ['GMMU'] = '0' # for usbgpu fast loading, noop for qcom
 from tinygrad.tensor import Tensor
 import time
 import pickle
 import numpy as np
 import cereal.messaging as messaging
 from cereal import car, log
-from pathlib import Path
 from cereal.messaging import PubMaster, SubMaster
 from msgq.visionipc import VisionIpcClient, VisionStreamType, VisionBuf
 from opendbc.car.car_helpers import get_demo_car_params
@@ -26,53 +20,50 @@ 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, WARP_INPUTS, POLICY_INPUTS
 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.common.file_chunker import read_file_chunked, get_manifest_path
 from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
+from openpilot.selfdrive.modeld.helpers import usbgpu_present, modeld_pkl_path, get_tg_input_devices
 
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
 
-
 PROCESS_NAME = "selfdrive.modeld.modeld"
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
 
-MODELS_DIR = Path(__file__).parent / 'models'
-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,
                           lat_action_t: float, long_action_t: float, v_ego: float) -> log.ModelDataV2.Action:
+  if 'action' not in model_output:
     plan = model_output['plan'][0]
     desired_accel, should_stop = get_accel_from_plan(plan[:,Plan.VELOCITY][:,0],
                                                      plan[:,Plan.ACCELERATION][:,0],
                                                      ModelConstants.T_IDXS,
                                                      action_t=long_action_t)
-    desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, LONG_SMOOTH_SECONDS)
-
     desired_curvature = get_curvature_from_plan(plan[:,Plan.T_FROM_CURRENT_EULER][:,2],
                                                 plan[:,Plan.ORIENTATION_RATE][:,2],
                                                 ModelConstants.T_IDXS,
                                                 v_ego,
                                                 lat_action_t)
-    if v_ego > MIN_LAT_CONTROL_SPEED:
-      desired_curvature = smooth_value(desired_curvature, prev_action.desiredCurvature, LAT_SMOOTH_SECONDS)
-    else:
-      desired_curvature = prev_action.desiredCurvature
+  else:
+    desired_accel = model_output['action'][0,1]
+    desired_curvature = model_output['action'][0,0] / (max(1.0, v_ego))**2
+    should_stop = (v_ego < 0.3 and desired_accel < 0.1)
+  desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, LONG_SMOOTH_SECONDS)
+  if v_ego > MIN_LAT_CONTROL_SPEED:
+    desired_curvature = smooth_value(desired_curvature, prev_action.desiredCurvature, LAT_SMOOTH_SECONDS)
+  else:
+    desired_curvature = prev_action.desiredCurvature
+
+  return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature),
+                                desiredAcceleration=float(desired_accel),
+                                shouldStop=bool(should_stop))
 
-    return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature),
-                                  desiredAcceleration=float(desired_accel),
-                                  shouldStop=bool(should_stop))
 
 class FrameMeta:
   frame_id: int = 0
@@ -83,176 +74,95 @@ 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, usbgpu: bool):
     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]
+    input_devices = get_tg_input_devices(PROCESS_NAME, usbgpu)
+    self.WARP_DEV, self.QUEUE_DEV = input_devices['WARP_DEV'], input_devices['QUEUE_DEV']
+    jits = pickle.loads(read_file_chunked(modeld_pkl_path(usbgpu)))
+    vision_metadata = jits['metadata']['vision']
+    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']
 
-    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]
+    policy_metadata = jits['metadata']['on_policy']
+    self.policy_input_shapes = policy_metadata['input_shapes']
+    self.policy_output_slices = policy_metadata['output_slices']
 
     self.prev_desire = np.zeros(ModelConstants.DESIRE_LEN, dtype=np.float32)
 
-    # policy inputs
-    self.numpy_inputs = {k: np.zeros(self.policy_input_shapes[k], dtype=np.float32) for k in self.policy_input_shapes}
-    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.frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
+    self.input_queues, self.npy = make_input_queues(self.vision_input_shapes, self.policy_input_shapes, self.frame_skip, device=self.QUEUE_DEV)
+    self.full_frames: dict[str, Tensor] = {}
+    self._blob_cache: dict[int, Tensor] = {}
     self.parser = Parser()
-    self.frame_buf_params : dict[str, tuple[int, int, int, int]] = {}
-    self.update_imgs = None
-    self.vision_run = pickle.loads(read_file_chunked(str(VISION_PKL_PATH)))
-    self.policy_run = pickle.loads(read_file_chunked(str(POLICY_PKL_PATH)))
+    self.frame_buf_params = {k: get_nv12_info(cam_w, cam_h) for k in ('img', 'big_img')}
+    self.run_policy = jits['run_policy']
+    self.warp_enqueue = jits[(cam_w,cam_h)]
 
   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()}
     return parsed_model_outputs
 
   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)
-
+          inputs: dict[str, np.ndarray], prepare_only: bool) -> dict[str, np.ndarray] | None:
     for key in bufs.keys():
-      ptr = bufs[key].data.ctypes.data
+      ptr = np.frombuffer(bufs[key].data, dtype=np.uint8).ctypes.data
       yuv_size = self.frame_buf_params[key][3]
       # There is a ringbuffer of imgs, just cache tensors pointing to all of them
       cache_key = (key, ptr)
       if cache_key not in self._blob_cache:
-        self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8')
+        self._blob_cache[cache_key] = Tensor.from_blob(ptr, (yuv_size,), dtype='uint8', device=self.WARP_DEV)
       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'],
-                           self.img_queues['big_img'], self.full_frames['big_img'], self.transforms['big_img'])
-    self.img_queues['img'], self.img_queues['big_img'] = out[0].realize(), out[2].realize()
-    vision_inputs = {'img': out[1], 'big_img': out[3]}
+    # Model decides when action is completed, so desire input is just a pulse triggered on rising edge
+    inputs['desire_pulse'][0] = 0
+    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['action_t'][:] = inputs['action_t']
+    self.npy['tfm'][:,:] = transforms['img'][:,:]
+    self.npy['big_tfm'][:,:] = transforms['big_img'][:,:]
+
+    img, big_img = self.warp_enqueue(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames['img'], big_frame=self.full_frames['big_img'])
 
     if prepare_only:
       return None
 
-    self.vision_output = self.vision_run(**vision_inputs).contiguous().realize().uop.base.buffer.numpy().flatten()
-    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(self.vision_output, self.vision_output_slices))
+    vision_output, on_policy_output = self.run_policy(
+      **{k: self.input_queues[k] for k in POLICY_INPUTS}, img=img, big_img=big_img
+    )
 
-    self.full_input_queues.enqueue({'features_buffer': vision_outputs_dict['hidden_state'], 'desire_pulse': new_desire})
-    for k in ['desire_pulse', 'features_buffer']:
-      self.numpy_inputs[k][:] = self.full_input_queues.get(k)[k]
-    self.numpy_inputs['traffic_convention'][:] = inputs['traffic_convention']
-
-    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))
+    vision_output = vision_output.numpy().flatten()
+    on_policy_output = on_policy_output.numpy().flatten()
+    vision_outputs_dict = self.parser.parse_vision_outputs(self.slice_outputs(vision_output, self.vision_output_slices))
+    policy_outputs_dict = self.parser.parse_policy_outputs(self.slice_outputs(on_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(), on_policy_output.copy()])
     return combined_outputs_dict
 
 
 def main(demo=False):
   cloudlog.warning("modeld init")
 
+  _present = usbgpu_present()
+  _compiled = os.path.isfile(get_manifest_path(modeld_pkl_path(usbgpu=True)))
+  USBGPU = _present and _compiled
+  params = Params()
+  params.put_bool("UsbGpuPresent", _present)
+  params.put_bool("UsbGpuCompiled", _compiled)
+
   if not USBGPU:
     # USB GPU currently saturates a core so can't do this yet,
     # 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)
@@ -276,6 +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, USBGPU)
+  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"])
@@ -296,7 +211,6 @@ def main(demo=False):
   meta_main = FrameMeta()
   meta_extra = FrameMeta()
 
-
   if demo:
     CP = get_demo_car_params()
   else:
@@ -380,9 +294,14 @@ def main(demo=False):
 
     bufs = {name: buf_extra if 'big' in name else buf_main for name in model.vision_input_names}
     transforms = {name: model_transform_extra if 'big' in name else model_transform_main for name in model.vision_input_names}
-    inputs:dict[str, np.ndarray] = {
+    frame_delay = DT_MDL # compensate for time passed since the frame was captured: current_time - timestamp_eof is 50ms on average
+    action_delay = DT_MDL / 2 # middle of the interval between model output (current state) and next frame (expected state)
+    lat_action_t = lat_delay + frame_delay + action_delay
+    long_action_t = long_delay + frame_delay + action_delay
+    inputs: dict[str, np.ndarray] = {
       'desire_pulse': vec_desire,
       'traffic_convention': traffic_convention,
+      'action_t': np.array([lat_action_t, long_action_t], dtype=np.float32),
     }
 
     mt1 = time.perf_counter()
@@ -396,9 +315,7 @@ def main(demo=False):
       posenet_send = messaging.new_message('cameraOdometry')
       mdv2sp_send = messaging.new_message('modelDataV2SP')
 
-      frame_delay = DT_MDL # compensate for time passed since the frame was captured: current_time - timestamp_eof is 50ms on average
-      action_delay = DT_MDL / 2 # middle of the interval between model output (current state) and next frame (expected state)
-      action = get_action_from_model(model_output, prev_action, lat_delay + frame_delay + action_delay, long_delay + frame_delay + action_delay, v_ego)
+      action = get_action_from_model(model_output, prev_action, lat_action_t, long_action_t, v_ego)
       prev_action = action
       fill_model_msg(drivingdata_send, modelv2_send, model_output, action,
                      publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id,

selfdrive/modeld/models/big_driving_on_policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:565e53c38dcd64c50dd3fe4d5ee1530213aeefd66c3f6b67ea6a72a32612a6bf
+size 14061419

selfdrive/modeld/models/big_driving_policy.onnx

@@ -1 +0,0 @@
-driving_policy.onnx

selfdrive/modeld/models/big_driving_vision.onnx

@@ -1 +0,0 @@
-driving_vision.onnx

selfdrive/modeld/models/big_driving_vision.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f0cab5033fe9e3bc5e174a2e790fa277f7d9fc44c65822d734064d2f899a9a0
+size 296203378