clean up some MORE

Update model_replay.py
rm
2026-09-10 00:34:55 +08:00 · 2026-09-09 08:58:19 -07:00 · 2026-09-09 00:56:07 -07:00 · 2026-09-09 00:24:09 -07:00 · 2026-09-09 00:08:41 -07:00 · 2026-09-09 00:05:23 -07:00
25 changed files with 777 additions and 572 deletions
@@ -1,11 +0,0 @@
-*       @sunnypilot/dev-internal
-/.github/ @devtekve @sunnyhaibin
-/release/ci/ @devtekve @sunnyhaibin
-/tinygrad_repo @devtekve @Discountchubbs
-/tinygrad/ @devtekve @Discountchubbs
-/selfdrive/controls/lib/longitudinal_planner.py @devtekve @Discountchubbs
-/selfdrive/controls/lib/longitudinal_mpc_lib/long_mpc.py @devtekve @Discountchubbs
-/selfdrive/modeld/ @devtekve @Discountchubbs
-/sunnypilot/model* @devtekve @Discountchubbs
-/sunnypilot/sunnylink/ @devtekve
-/system/athena/ @devtekve
@@ -78,6 +78,7 @@ jobs:
      - name: Get next recompiled dir number
        id: create-recompiled-dir
        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
          HF_REPO: ${{ github.event.inputs.hf_repo }}
        run: |
          pip install huggingface_hub
@@ -30,6 +30,7 @@ jobs:
    runs-on: ubuntu-24.04
    outputs:
      model_name: ${{ steps.resolve.outputs.model_name }}
+      safe_model_name: ${{ steps.resolve.outputs.safe_model_name }}
      onnx_ref: ${{ steps.resolve.outputs.onnx_ref }}
      onnx_path: ${{ steps.resolve.outputs.onnx_path }}
      hf_defaults_path: ${{ steps.resolve.outputs.hf_defaults_path }}
@@ -64,7 +65,9 @@ jobs:
            exit 1
          fi

+          SAFE_NAME="${NAME// /-}"
          echo "model_name=${NAME}" >> $GITHUB_OUTPUT
+          echo "safe_model_name=${SAFE_NAME}" >> $GITHUB_OUTPUT
          echo "onnx_ref=${ONNX_REF}" >> $GITHUB_OUTPUT
          echo "onnx_path=${ONNX_PATH}" >> $GITHUB_OUTPUT
          echo "hf_defaults_path=${HF_DEFAULTS_PATH}" >> $GITHUB_OUTPUT
@@ -135,7 +138,7 @@ jobs:

      - name: Prepare output
        env:
-          MODEL_NAME: ${{ needs.resolve.outputs.model_name }}
+          MODEL_NAME: ${{ needs.resolve.outputs.safe_model_name }}
        run: |
          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
          export PYTHONPATH=${{ github.workspace }}
@@ -158,13 +161,13 @@ jobs:
      - name: Upload small model artifact
        uses: actions/upload-artifact@v4
        with:
-          name: model-${{ needs.resolve.outputs.model_name }}-${{ github.run_number }}
+          name: model-${{ needs.resolve.outputs.safe_model_name }}-${{ github.run_number }}
          path: ${{ github.workspace }}/small_output/

      - name: Upload artifact name file
        uses: actions/upload-artifact@v4
        with:
-          name: artifact-name-${{ needs.resolve.outputs.model_name }}
+          name: artifact-name-${{ needs.resolve.outputs.safe_model_name }}
          path: ${{ github.workspace }}/small_output/artifact_name.txt

      - name: Re-enable powersave
@@ -254,7 +257,7 @@ jobs:

      - name: Prepare output
        env:
-          MODEL_NAME: ${{ needs.resolve.outputs.model_name }}
+          MODEL_NAME: ${{ needs.resolve.outputs.safe_model_name }}
        run: |
          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
          export PYTHONPATH=${{ github.workspace }}
@@ -277,13 +280,13 @@ jobs:
      - name: Upload big model artifact
        uses: actions/upload-artifact@v4
        with:
-          name: model-${{ needs.resolve.outputs.model_name }}-${{ github.run_number }}
+          name: model-${{ needs.resolve.outputs.safe_model_name }}-${{ github.run_number }}
          path: ${{ github.workspace }}/big_output/

      - name: Upload artifact name file
        uses: actions/upload-artifact@v4
        with:
-          name: artifact-name-${{ needs.resolve.outputs.model_name }}
+          name: artifact-name-${{ needs.resolve.outputs.safe_model_name }}
          path: ${{ github.workspace }}/big_output/artifact_name.txt

      - name: Re-enable powersave
@@ -318,7 +321,7 @@ jobs:
        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        uses: actions/download-artifact@v4
        with:
-          name: artifact-name-${{ needs.resolve.outputs.model_name }}
+          name: artifact-name-${{ needs.resolve.outputs.safe_model_name }}
          path: artifact_name

      - name: Read artifact name
@@ -338,7 +341,7 @@ jobs:
      - name: Upload model to HF
        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        env:
-          HF_OIDC_RESOURCE: datasets/${{ env.HF_REPO }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
          ARTIFACT_NAME: ${{ steps.artifact.outputs.artifact_name }}
        run: |
          rm -f output/artifact_name.txt
@@ -364,7 +367,7 @@ jobs:
      - name: Generate DM metadata and upload to HF
        if: ${{ inputs.target == 'dm' }}
        env:
-          HF_OIDC_RESOURCE: datasets/${{ env.HF_REPO }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          export PYTHONPATH=$(pwd)
          python3 -c "
@@ -481,11 +484,29 @@ jobs:
          print(f'Chunked {pkl} into {len(targets)} chunks')
          "

+      - name: Compile DM warp
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH="${PYTHONPATH}:${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}"
+
+          TG_FLAGS="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+          MODEL_DIR="${{ github.workspace }}/openpilot/selfdrive/modeld"
+          DM_SIZE=$(python3 -c "from openpilot.common.transformations.model import DM_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
+
+          for res in $(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')"); do
+            WARP_PKL="${MODEL_DIR}/models/dm_warp_${res}_tinygrad.pkl"
+            taskset -c 7 env ${TG_FLAGS} python3 ${MODEL_DIR}/compile_dm_warp.py \
+              --camera-resolution ${res} \
+              --warp-to ${DM_SIZE} \
+              --output ${WARP_PKL}
+          done
+
      - name: Prepare DM output
        run: |
          mkdir -p dm_output
          cp ${{ github.workspace }}/${{ env.DM_PKL }}.chunk* dm_output/
          cp ${{ github.workspace }}/${{ env.DM_PKL }}.chunkmanifest dm_output/
+          cp ${{ github.workspace }}/openpilot/selfdrive/modeld/models/dm_warp_* dm_output/

      - name: Upload DM artifact
        uses: actions/upload-artifact@v4
@@ -146,7 +146,7 @@ jobs:

      - name: Validate hf_repo and JSON version
        env:
-          HF_OIDC_RESOURCE: datasets/${{ inputs.hf_repo }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
        run: |
          if [ ! -f "$JSON_FILE" ]; then
            echo "JSON file $JSON_FILE does not exist!"
@@ -155,13 +155,8 @@ jobs:
          python3 -c "
          import sys
          from huggingface_hub import HfApi
-          try:
-              api = HfApi()
-              api.repo_info(repo_id=sys.argv[1], repo_type='dataset')
-              print(f'Success: Repo {sys.argv[1]} exists.')
-          except Exception as e:
-              print('HF validation failed:', e)
-              sys.exit(1)
+          HfApi().repo_info(repo_id=sys.argv[1], repo_type='dataset')
+          print(f'Success: Repo {sys.argv[1]} exists.')
          " "${{ inputs.hf_repo }}"

      - name: Download artifact name file
@@ -192,7 +187,7 @@ jobs:

      - name: Upload to Hugging Face
        env:
-          HF_OIDC_RESOURCE: datasets/${{ inputs.hf_repo }}
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
          ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
        run: |
          hf upload ${{ inputs.hf_repo }} \
@@ -46,6 +46,13 @@ runs:
            printf '%s\t%s\n' "$ENCODED_URL" "${DEST_DIR}/${CANONICAL}.chunk${CHUNK_IDX}" >> "$DOWNLOAD_LIST"
          done < <(echo "$ARTIFACT" | jq -r '.chunks[].file_name')
          echo "$NUM_CHUNKS" > "${DEST_DIR}/${CANONICAL}.chunkmanifest"
+
+          if [ "$CANONICAL" = "dmonitoring_model_tinygrad.pkl" ]; then
+            for warp in dm_warp_1928x1208_tinygrad.pkl dm_warp_1344x760_tinygrad.pkl; do
+              ENCODED_URL=$(python3 -c "import urllib.parse; print(urllib.parse.quote('${BASE_URL}/${warp}', safe=':/'))")
+              printf '%s\t%s\n' "$ENCODED_URL" "${DEST_DIR}/${warp}" >> "$DOWNLOAD_LIST"
+            done
+          fi
        }

        echo "$MODELS_JSON" | jq -c '.[]' | while IFS= read -r model; do
@@ -0,0 +1,148 @@
+name: Test Stock vs Sunnypilot Model Equivalence
+
+on:
+  workflow_dispatch:
+    inputs:
+      model_ref:
+        description: 'Upstream openpilot commit ref'
+        required: false
+        default: ''
+  pull_request:
+    paths:
+      - 'openpilot/selfdrive/modeld/**'
+      - 'openpilot/sunnypilot/modeld_v2/**'
+
+jobs:
+  test_stock_parity:
+    name: Compare Stock vs Sunnypilot Model Replay
+    runs-on: macos-latest
+
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - run: ./tools/op.sh setup
+      - run: scons -j$(nproc 2>/dev/null || sysctl -n hw.logicalcpu) openpilot/cereal msgq_repo openpilot/common
+      - name: Fetch Big Model ONNX
+        run: |
+          mkdir -p /tmp/onnx_models
+          if [ -n "${{ inputs.model_ref }}" ]; then
+            echo "Fetching ONNX from upstream openpilot ref ${{ inputs.model_ref }}..."
+            git clone --depth 1 https://github.com/commaai/openpilot.git /tmp/upstream_openpilot
+            cd /tmp/upstream_openpilot
+            git fetch --depth 1 origin ${{ inputs.model_ref }}
+            git checkout ${{ inputs.model_ref }}
+            git lfs pull -I "**/selfdrive/modeld/models/big_driving_supercombo.onnx"
+            find . -name "big_driving_supercombo.onnx" -exec cp {} /tmp/onnx_models/ \;
+          else
+            cp openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx /tmp/onnx_models/
+          fi
+
+      - name: Compile models
+        env:
+          DEV: "CPU"
+          JIT_BATCH_SIZE: "0"
+        run: |
+          BIG_ONNX="/tmp/onnx_models/big_driving_supercombo.onnx"
+          MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
+          CAMERA_RESOLUTIONS=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
+
+          python3 openpilot/selfdrive/modeld/compile_modeld.py \
+            --onnx "$BIG_ONNX" \
+            --model-size "$MODEL_SIZE" \
+            --camera-resolutions $CAMERA_RESOLUTIONS \
+            --output /tmp/stock_model.pkl \
+            --frame-skip 4 \
+            --benchmark-runs 1 &
+
+          python3 openpilot/sunnypilot/modeld_v2/compile_modeld.py \
+            --model-type supercombo \
+            --supercombo-onnx "$BIG_ONNX" \
+            --model-size "$MODEL_SIZE" \
+            --camera-resolutions $CAMERA_RESOLUTIONS \
+            --output /tmp/sunnypilot_model.pkl \
+            --frame-skip 4 \
+            --benchmark-runs 1 &
+          wait
+
+      - name: Run model replay
+        env:
+          DEV: "CPU"
+        run: |
+          python3 openpilot/sunnypilot/modeld_v2/model_replay.py \
+            --sunnypilot-model /tmp/sunnypilot_model.pkl \
+            --stock-model /tmp/stock_model.pkl \
+            --frames 20 \
+            --plot-dir /tmp/replay_plots
+
+      - name: Upload Replay Plots
+        uses: actions/upload-artifact@v4
+        if: always()
+        continue-on-error: true
+        with:
+          name: model_replay_plots_${{ github.event.number || github.sha }}
+          path: /tmp/replay_plots
+
+      - name: Checkout ci-artifacts
+        if: github.repository == 'sunnypilot/sunnypilot' && github.event_name == 'pull_request'
+        uses: actions/checkout@v4
+        with:
+          repository: sunnypilot/ci-artifacts
+          ssh-key: ${{ secrets.CI_ARTIFACTS_DEPLOY_KEY }}
+          path: ${{ github.workspace }}/ci-artifacts
+
+      - name: Push plots to ci-artifacts
+        if: github.repository == 'sunnypilot/sunnypilot' && github.event_name == 'pull_request'
+        working-directory: ${{ github.workspace }}/ci-artifacts
+        run: |
+          git config user.name "GitHub Actions Bot"
+          git config user.email "<>"
+          BRANCH="model_replay_pr_${{ github.event.number }}"
+          git fetch origin $BRANCH || true
+          git checkout $BRANCH 2>/dev/null || git checkout --orphan $BRANCH
+          rm -rf plots && mkdir -p plots
+          cp /tmp/replay_plots/*.png plots/
+          echo "${{ github.sha }}" > ref_commit
+          git add plots ref_commit
+          git commit -m "Model replay plots for PR #${{ github.event.number }}@${{ github.sha }}" || echo "No changes to commit"
+          git push origin $BRANCH --force
+
+      - name: Comment Model Replay Report on PR
+        if: github.repository == 'sunnypilot/sunnypilot' && github.event_name == 'pull_request'
+        uses: actions/github-script@v7
+        with:
+          script: |
+            const fs = require('fs');
+            const prNumber = context.payload.pull_request.number;
+            const branch = `model_replay_pr_${prNumber}`;
+            const baseUrl = `https://raw.githubusercontent.com/sunnypilot/ci-artifacts/refs/heads/${branch}/plots`;
+
+            const priorityPlots = ['desiredCurvature.png', 'desiredAcceleration.png', 'velocity.x.png', 'leadsV3.x.png', 'execution_timings.png'];
+            const allFiles = fs.readdirSync('/tmp/replay_plots').filter(f => f.endsWith('.png'));
+            const orderedFiles = [
+              ...priorityPlots.filter(f => allFiles.includes(f)),
+              ...allFiles.filter(f => !priorityPlots.includes(f)).sort()
+            ];
+
+            let table = '<table>';
+            for (let i = 0; i < orderedFiles.length; i += 2) {
+              table += '<tr>';
+              table += `<td><img src="${baseUrl}/${orderedFiles[i]}" alt="${orderedFiles[i]}"><br><b>${orderedFiles[i].replace('.png', '')}</b></td>`;
+              if (i + 1 < orderedFiles.length) {
+                table += `<td><img src="${baseUrl}/${orderedFiles[i+1]}" alt="${orderedFiles[i+1]}"><br><b>${orderedFiles[i+1].replace('.png', '')}</b></td>`;
+              } else {
+                table += '<td></td>';
+              }
+              table += '</tr>';
+            }
+            table += '</table>';
+
+            const body = `### Model Replay Parity Report for PR #${prNumber} (@${context.sha.substring(0, 7)})\n\n` +
+              `<details><summary>All Model Replay Plots</summary>\n\n${table}\n\n</details>`;
+
+            await github.rest.issues.createComment({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: prNumber,
+              body: body
+            });
@@ -188,7 +188,7 @@ jobs:
          if [ "${{ inputs.target_hardware }}" == "chestnut" ]; then
            echo "CHESTNUT build"
            export CHESTNUT=1
-            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2"
+            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2 TC_OCCUPANCY_OPT=1"
            OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
          else
            echo "QCOM build"
@@ -216,6 +216,9 @@ jobs:
    needs: [ prepare_strategy ]
    runs-on: ubuntu-24.04
    if: ${{ needs.prepare_strategy.outputs.include_big_model == 'true' }}
+    concurrency:
+      group: prepare-chestnut
+      cancel-in-progress: false
    outputs:
      onnx_sha256: ${{ steps.resolve.outputs.onnx_sha256 }}
    env:
@@ -228,8 +231,10 @@ jobs:
        run: |
          REF="${{ github.head_ref || github.ref_name }}"

-          ONNX_HASH=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx?ref=${REF}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
+          BLOB_SHA=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx?ref=${REF}" --jq '.sha')
+          ONNX_HASH=$(gh api "repos/${GH_REPO}/git/blobs/${BLOB_SHA}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
          echo "ONNX hash: $ONNX_HASH"
+          [ -n "$ONNX_HASH" ] || { echo "::error::Failed to extract ONNX hash"; exit 1; }
          echo "onnx_sha256=$ONNX_HASH" >> $GITHUB_OUTPUT

          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
@@ -238,7 +243,7 @@ jobs:
          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"

          check_defaults() {
-            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
+            DEFAULTS=$(curl -fsSL "${JSON_URL}?t=$(date +%s)" 2>/dev/null) || return 1
            TINYGRAD_MATCH=$(echo "$DEFAULTS" | jq -r --arg ref "$TINYGRAD_REF" '.tinygrad_ref == $ref' 2>/dev/null)
            [ "$TINYGRAD_MATCH" = "true" ] || return 1
            BUNDLE=$(echo "$DEFAULTS" | jq --arg hash "$ONNX_HASH" '.bundles[] | select(.onnx_sha256 == $hash)' 2>/dev/null)
@@ -252,18 +257,35 @@ jobs:

          echo "No matching model on HF — dispatching build"
          gh workflow run build-default-models.yaml --ref "$REF" -f target=big
+          sleep 10

-          echo "Polling HF for big model availability..."
+          BUILD_RUN_ID=$(gh run list --workflow build-default-models.yaml --branch "$REF" --limit 1 --json databaseId --jq '.[0].databaseId')
+          echo "Dispatched build run: $BUILD_RUN_ID"
+
+          echo "Waiting for build run to complete..."
          for i in $(seq 1 90); do
            sleep 30
-            if check_defaults; then
-              echo "Big model available on HF after $((i * 30))s"
-              exit 0
+            STATUS=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.status')
+            CONCLUSION=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.conclusion')
+            echo "Poll $i/90: status=$STATUS conclusion=$CONCLUSION"
+            if [ "$STATUS" = "completed" ]; then
+              if [ "$CONCLUSION" = "success" ]; then
+                echo "Build run succeeded, verifying HF..."
+                sleep 10
+                if check_defaults; then
+                  echo "Big model verified on HF"
+                  exit 0
+                fi
+                echo "::error::Build succeeded but model not found on HF"
+                exit 1
+              else
+                echo "::error::Build run failed with conclusion=$CONCLUSION"
+                exit 1
+              fi
            fi
-            echo "Poll $i/90: not yet available"
          done

-          echo "::error::Big model not available on HF after 45 minutes"
+          echo "::error::Build run did not complete within 45 minutes"
          exit 1
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -277,6 +299,9 @@ jobs:
  prepare_small_model:
    needs: [ prepare_strategy ]
    runs-on: ubuntu-24.04
+    concurrency:
+      group: prepare-small-model
+      cancel-in-progress: false
    outputs:
      driving_onnx_sha256: ${{ steps.resolve.outputs.driving_onnx_sha256 }}
    env:
@@ -289,8 +314,10 @@ jobs:
        run: |
          REF="${{ github.head_ref || github.ref_name }}"

-          DRIVING_HASH=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/driving_supercombo.onnx?ref=${REF}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
+          BLOB_SHA=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/driving_supercombo.onnx?ref=${REF}" --jq '.sha')
+          DRIVING_HASH=$(gh api "repos/${GH_REPO}/git/blobs/${BLOB_SHA}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
          echo "Driving ONNX hash: $DRIVING_HASH"
+          [ -n "$DRIVING_HASH" ] || { echo "::error::Failed to extract driving ONNX hash"; exit 1; }
          echo "driving_onnx_sha256=$DRIVING_HASH" >> $GITHUB_OUTPUT

          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
@@ -299,7 +326,7 @@ jobs:
          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"

          check_defaults() {
-            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
+            DEFAULTS=$(curl -fsSL "${JSON_URL}?t=$(date +%s)" 2>/dev/null) || return 1
            TINYGRAD_MATCH=$(echo "$DEFAULTS" | jq -r --arg ref "$TINYGRAD_REF" '.tinygrad_ref == $ref' 2>/dev/null)
            [ "$TINYGRAD_MATCH" = "true" ] || return 1
            DRIVING=$(echo "$DEFAULTS" | jq --arg hash "$DRIVING_HASH" '.bundles[] | select(.onnx_sha256 == $hash)' 2>/dev/null)
@@ -313,18 +340,35 @@ jobs:

          echo "No matching model on HF — dispatching build"
          gh workflow run build-default-models.yaml --ref "$REF" -f target=small
+          sleep 10

-          echo "Polling HF for model availability..."
+          BUILD_RUN_ID=$(gh run list --workflow build-default-models.yaml --branch "$REF" --limit 1 --json databaseId --jq '.[0].databaseId')
+          echo "Dispatched build run: $BUILD_RUN_ID"
+
+          echo "Waiting for build run to complete..."
          for i in $(seq 1 60); do
            sleep 30
-            if check_defaults; then
-              echo "Model available on HF after $((i * 30))s"
-              exit 0
+            STATUS=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.status')
+            CONCLUSION=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.conclusion')
+            echo "Poll $i/60: status=$STATUS conclusion=$CONCLUSION"
+            if [ "$STATUS" = "completed" ]; then
+              if [ "$CONCLUSION" = "success" ]; then
+                echo "Build run succeeded, verifying HF..."
+                sleep 10
+                if check_defaults; then
+                  echo "Small model verified on HF"
+                  exit 0
+                fi
+                echo "::error::Build succeeded but model not found on HF"
+                exit 1
+              else
+                echo "::error::Build run failed with conclusion=$CONCLUSION"
+                exit 1
+              fi
            fi
-            echo "Poll $i/60: not yet available"
          done

-          echo "::error::Small driving model not available on HF after 30 minutes"
+          echo "::error::Small model build did not complete within 30 minutes"
          exit 1
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -338,6 +382,9 @@ jobs:
  prepare_dm_model:
    needs: [ prepare_strategy ]
    runs-on: ubuntu-24.04
+    concurrency:
+      group: prepare-dm-model
+      cancel-in-progress: false
    outputs:
      dm_onnx_sha256: ${{ steps.resolve.outputs.dm_onnx_sha256 }}
    env:
@@ -350,8 +397,10 @@ jobs:
        run: |
          REF="${{ github.head_ref || github.ref_name }}"

-          DM_HASH=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/dmonitoring_model.onnx?ref=${REF}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
+          BLOB_SHA=$(gh api "repos/${GH_REPO}/contents/openpilot/selfdrive/modeld/models/dmonitoring_model.onnx?ref=${REF}" --jq '.sha')
+          DM_HASH=$(gh api "repos/${GH_REPO}/git/blobs/${BLOB_SHA}" --jq '.content' | base64 -d | grep '^oid sha256:' | cut -d: -f2)
          echo "DM ONNX hash: $DM_HASH"
+          [ -n "$DM_HASH" ] || { echo "::error::Failed to extract DM ONNX hash"; exit 1; }
          echo "dm_onnx_sha256=$DM_HASH" >> $GITHUB_OUTPUT

          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
@@ -360,7 +409,7 @@ jobs:
          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"

          check_defaults() {
-            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
+            DEFAULTS=$(curl -fsSL "${JSON_URL}?t=$(date +%s)" 2>/dev/null) || return 1
            TINYGRAD_MATCH=$(echo "$DEFAULTS" | jq -r --arg ref "$TINYGRAD_REF" '.tinygrad_ref == $ref' 2>/dev/null)
            [ "$TINYGRAD_MATCH" = "true" ] || return 1
            DM=$(echo "$DEFAULTS" | jq --arg hash "$DM_HASH" '.bundles[] | select(.onnx_sha256 == $hash)' 2>/dev/null)
@@ -374,18 +423,35 @@ jobs:

          echo "No matching DM model on HF — dispatching build"
          gh workflow run build-default-models.yaml --ref "$REF" -f target=dm
+          sleep 10

-          echo "Polling HF for DM model availability..."
+          BUILD_RUN_ID=$(gh run list --workflow build-default-models.yaml --branch "$REF" --limit 1 --json databaseId --jq '.[0].databaseId')
+          echo "Dispatched build run: $BUILD_RUN_ID"
+
+          echo "Waiting for build run to complete..."
          for i in $(seq 1 60); do
            sleep 30
-            if check_defaults; then
-              echo "DM model available on HF after $((i * 30))s"
-              exit 0
+            STATUS=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.status')
+            CONCLUSION=$(gh api "repos/${GH_REPO}/actions/runs/${BUILD_RUN_ID}" --jq '.conclusion')
+            echo "Poll $i/60: status=$STATUS conclusion=$CONCLUSION"
+            if [ "$STATUS" = "completed" ]; then
+              if [ "$CONCLUSION" = "success" ]; then
+                echo "Build run succeeded, verifying HF..."
+                sleep 10
+                if check_defaults; then
+                  echo "DM model verified on HF"
+                  exit 0
+                fi
+                echo "::error::Build succeeded but DM model not found on HF"
+                exit 1
+              else
+                echo "::error::Build run failed with conclusion=$CONCLUSION"
+                exit 1
+              fi
            fi
-            echo "Poll $i/60: not yet available"
          done

-          echo "::error::DM model not available on HF after 30 minutes"
+          echo "::error::DM model build did not complete within 30 minutes"
          exit 1
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -1,5 +0,0 @@
-from pathlib import Path
-
-MODEL_PATH = Path(__file__).parent / 'models/supercombo.onnx'
-MODEL_PKL_PATH = Path(__file__).parent / 'models/supercombo_tinygrad.pkl'
-METADATA_PATH = Path(__file__).parent / 'models/supercombo_metadata.pkl'
@@ -32,7 +32,7 @@ def _patch_tinygrad_fetch_fw():
  helpers.fetch_fw = fetch_fw
 _patch_tinygrad_fetch_fw()

-from openpilot.selfdrive.modeld.compile_modeld import NV12Frame, make_frame_prepare, sample_desire, sample_skip, shift_and_sample
+import openpilot.selfdrive.modeld.compile_modeld as stock
 from tinygrad import dtypes
 from tinygrad.device import Device
 from tinygrad.engine.jit import TinyJit
@@ -41,8 +41,7 @@ from tinygrad.tensor import Tensor
 MODEL_TYPES = ('vision_policy', 'supercombo', 'vision_multi_policy')
 WARP_INPUTS = ['tfm', 'big_tfm']
 POLICY_INPUTS = ['img_q', 'big_img_q', 'feat_q', 'desire_q', 'packed_npy_inputs']
-WARP_DEV = os.getenv('WARP_DEV')
-
+nv12_copy_size = stock.nv12_copy_size

 def _detect_desire_key(shapes: dict) -> str | None:
  return next((key for key in shapes if key.startswith('desire')), None)
@@ -139,7 +138,7 @@ def make_supercombo_input_queues(input_shapes: dict, frame_skip: int,
  return generate_queues_and_npy(input_shapes, frame_skip, device, is_supercombo=True)


-def make_random_images(keys, shape, device):
+def make_random_images(keys, shape, device, rng=None):
  return {k: Tensor.randint(shape, low=0, high=256, dtype=dtypes.uint8, device=device).realize() for k in keys}


@@ -152,24 +151,9 @@ def make_warp_queues(device=Device.DEFAULT):
  return queues, npy


-def make_warp(nv12: NV12Frame, model_w: int, model_h: int):
-  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
-  WARP_DEV = os.getenv('WARP_DEV', Device.DEFAULT)
-
-  def warp(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)
-    warped_big_frame = frame_prepare(big_frame, big_tfm).unsqueeze(0)
-    return Tensor.cat(warped_frame, warped_big_frame)
-  return warp
-
-
 def make_run_policy(vision_runner, policy_runners: list, features_slice: slice, frame_skip: int, input_shapes: dict):
-  sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
-  sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
+  sample_skip_fn = partial(stock.sample_skip, frame_skip=frame_skip)
+  sample_desire_fn = partial(stock.sample_desire, frame_skip=frame_skip)

  desire_key = _detect_desire_key(input_shapes)
  road_key, wide_key = _detect_vision_keys(input_shapes)
@@ -186,14 +170,14 @@ def make_run_policy(vision_runner, policy_runners: list, features_slice: slice,
    warped_dev = warped.to(Device.DEFAULT)
    Tensor.realize(packed_npy_inputs_dev, warped_dev)

-    img = shift_and_sample(img_q, warped_dev[0:1], sample_skip_fn)
-    big_img = shift_and_sample(big_img_q, warped_dev[1:2], sample_skip_fn)
+    img = stock.shift_and_sample(img_q, warped_dev[0:1], sample_skip_fn)
+    big_img = stock.shift_and_sample(big_img_q, warped_dev[1:2], sample_skip_fn)

    unpacked_tensors = [tensor.reshape(shape) for tensor, shape in zip(packed_npy_inputs_dev.split(npy_sizes), npy_shapes.values(), strict=True)]
    unpacked_dict = dict(zip(npy_shapes.keys(), unpacked_tensors, strict=True))

    desire_dev = unpacked_dict['desire']
-    desire_buf = shift_and_sample(desire_q, desire_dev.reshape(1, 1, -1), sample_desire_fn)
+    desire_buf = stock.shift_and_sample(desire_q, desire_dev.reshape(1, 1, -1), sample_desire_fn)

    inputs = {desire_key: desire_buf}
    for key, tensor_val in unpacked_dict.items():
@@ -202,13 +186,13 @@ def make_run_policy(vision_runner, policy_runners: list, features_slice: slice,

    if 'prev_feat' in unpacked_dict:
      prev_feat_dev = unpacked_dict['prev_feat']
-      inputs['features_buffer'] = shift_and_sample(feat_q, prev_feat_dev.reshape(1, 1, -1), sample_skip_fn).reshape(input_shapes['features_buffer'])
+      inputs['features_buffer'] = stock.shift_and_sample(feat_q, prev_feat_dev.reshape(1, 1, -1), sample_skip_fn).reshape(input_shapes['features_buffer'])

    if vision_runner:
      vision_out_cast = next(iter(vision_runner({road_key: img, wide_key: big_img}).values())).cast('float32').realize()
      if 'features_buffer' not in inputs:
        new_feat = vision_out_cast[:, features_slice].reshape(1, -1).unsqueeze(0)
-        inputs['features_buffer'] = shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
+        inputs['features_buffer'] = stock.shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
      policy_outs = [next(iter(pol_runner(inputs).values())).cast('float32').realize() for pol_runner in policy_runners]
      return (vision_out_cast, *policy_outs) if len(policy_outs) > 1 else (vision_out_cast, policy_outs[0])

@@ -219,27 +203,28 @@ def make_run_policy(vision_runner, policy_runners: list, features_slice: slice,
    policy_out = next(iter(policy_runners[0](inputs).values())).cast('float32').realize()
    if 'features_buffer' not in inputs and features_slice is not None:
      new_feat = policy_out[:, features_slice].reshape(1, -1).unsqueeze(0)
-      shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
+      stock.shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
    return policy_out

  return run_policy


-def compile_jit(jit, make_random_inputs, input_keys, make_queues):
+def compile_jit(jit, input_keys, make_queues, make_random_inputs=None, benchmark_runs: int = 1):
  SEED = 42
-  def random_inputs_run(fn, seed, test_val=None, test_buffers=None, expect_match=True):
-    input_queues, npy = make_queues(Device.DEFAULT)
+  def random_inputs_run(fn, seed, n_runs, test_val=None, test_buffers=None, expect_match=True):
+    queues_res = make_queues(Device.DEFAULT)
+    input_queues, npy = queues_res[0], queues_res[1]
+    frame_views = queues_res[2] if len(queues_res) > 2 else {}
    rng = np.random.default_rng(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[:] = rng.standard_normal(v.shape).astype(v.dtype)
+      for v in frame_views.values():
+        v[:] = rng.integers(0, 256, size=v.shape, dtype=np.uint8)
      Device.default.synchronize()
-      random_inputs = make_random_inputs()
+      random_inputs = make_random_inputs(rng=rng) if make_random_inputs is not None else {}
      st = time.perf_counter()
      outs = fn(**{k: input_queues[k] for k in input_keys if k in input_queues}, **random_inputs)
      mt = time.perf_counter()
@@ -260,14 +245,15 @@ def compile_jit(jit, make_random_inputs, input_keys, make_queues):
    return val, buffers

  print('capture + replay')
-  test_val, test_buffers = random_inputs_run(jit, SEED)
-  print('pickle round trip')
+  test_val, test_buffers = random_inputs_run(jit, SEED, 3)
+  print(f'pickle round trip ({benchmark_runs} runs per seed)')
  with tempfile.TemporaryFile(dir=".") as f:
    dump_oob(jit, f)
    f.seek(0)
-    deserialized_jit = load_oob(f)
-  random_inputs_run(deserialized_jit, SEED, test_val=test_val, test_buffers=test_buffers)
-  return deserialized_jit
+    loaded_jit = load_oob(f)
+  random_inputs_run(loaded_jit, SEED, benchmark_runs, test_val, test_buffers, expect_match=True)
+  random_inputs_run(loaded_jit, SEED+1, benchmark_runs, test_val, test_buffers, expect_match=False)
+  return jit


 def _parse_size(size_str: str) -> tuple[int, int]:
@@ -317,6 +303,7 @@ if __name__ == "__main__":
  parser.add_argument('--model-size', type=_parse_size, required=True, help='model input WxH')
  parser.add_argument('--camera-resolutions', type=_parse_size, nargs='+', required=True)
  parser.add_argument('--frame-skip', type=int, default=None, help='frame skip value (auto-derived if not provided)')
+  parser.add_argument('--benchmark-runs', type=int, default=1, help='benchmark runs')
  parser.add_argument('--output', required=True)

  parser.add_argument('--vision-onnx', help='vision ONNX (for split models)')
@@ -335,48 +322,64 @@ if __name__ == "__main__":
  args.on_policy_onnx = read_file_chunked_to_disk(args.on_policy_onnx)
  args.supercombo_onnx = read_file_chunked_to_disk(args.supercombo_onnx)

-  vision_runner = OnnxRunner(args.vision_onnx) if args.vision_onnx else None
-
-  if args.model_type == 'vision_policy':
-    assert vision_runner and args.policy_onnx
-    policy_runners = [OnnxRunner(args.policy_onnx)]
-    output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx), 'policy': make_metadata_dict(args.policy_onnx)}
-  elif args.model_type == 'supercombo':
+  if args.model_type == 'supercombo':
    assert args.supercombo_onnx
-    policy_runners = [OnnxRunner(args.supercombo_onnx)]
-    output_data['metadata'] = {'model': make_metadata_dict(args.supercombo_onnx)}
-  elif args.model_type == 'vision_multi_policy':
-    assert vision_runner
-    policy_runners, policy_names = _load_policy_runners(args)
-    output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx)}
-    for name in policy_names:
-      runner_arg = getattr(args, f"{name}_onnx")
-      output_data['metadata'][name] = make_metadata_dict(runner_arg)
+    model_metadata = make_metadata_dict(args.supercombo_onnx)
+    output_data['metadata'] = {'model': model_metadata, **model_metadata}
+    output_data['input_devices'] = {'model': Device.DEFAULT}
+    output_data['run_model'] = {}
+    derived_frame_skip = args.frame_skip or derive_frame_skip({}, model_metadata['input_shapes'])
+    model_runner = OnnxRunner(args.supercombo_onnx)
+    run_policy = stock.make_run_policy(model_runner, model_metadata, derived_frame_skip)
+    for cam_w, cam_h in args.camera_resolutions:
+      print(f"Compiling unified run_model JIT for {cam_w}x{cam_h}...")
+      nv12 = stock.NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+      frame_copy_size = stock.nv12_copy_size(nv12.stride, nv12.y_height, nv12.uv_height)
+      make_model_queues = partial(stock.make_input_queues, model_metadata['input_shapes'], derived_frame_skip,
+                                  frame_copy_size=frame_copy_size)
+      warp = stock.make_warp(nv12, model_w, model_h)
+      run_model_jit = TinyJit(stock.make_run_model(warp, run_policy, model_metadata, frame_copy_size), prune=True)
+      output_data['run_model'][(cam_w, cam_h)] = compile_jit(run_model_jit, stock.MODELD_INPUTS, make_model_queues, benchmark_runs=args.benchmark_runs)
+  else:
+    vision_runner = OnnxRunner(args.vision_onnx) if args.vision_onnx else None
+    if args.model_type == 'vision_policy':
+      assert vision_runner and args.policy_onnx
+      policy_runners = [OnnxRunner(args.policy_onnx)]
+      output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx), 'policy': make_metadata_dict(args.policy_onnx)}
+    elif args.model_type == 'vision_multi_policy':
+      assert vision_runner
+      policy_runners, policy_names = _load_policy_runners(args)
+      output_data['metadata'] = {'vision': make_metadata_dict(args.vision_onnx)}
+      for name in policy_names:
+        runner_arg = getattr(args, f"{name}_onnx")
+        output_data['metadata'][name] = make_metadata_dict(runner_arg)

-  policy_keys = [key for key in output_data['metadata'].keys() if key != 'vision']
-  first_policy_meta = output_data['metadata'][policy_keys[0]] if policy_keys else {}
-  vision_meta = output_data['metadata'].get('vision', {})
+    policy_keys = [key for key in output_data['metadata'].keys() if key != 'vision']
+    first_policy_meta = output_data['metadata'][policy_keys[0]] if policy_keys else {}
+    vision_meta = output_data['metadata'].get('vision', {})

-  derived_frame_skip = args.frame_skip or derive_frame_skip(vision_meta.get('input_shapes', {}), first_policy_meta.get('input_shapes', {}))
-  all_shapes = {key: value for meta in output_data['metadata'].values() for key, value in meta['input_shapes'].items()}
-  feat_meta = output_data['metadata'].get('vision') or output_data['metadata'].get('model') or output_data['metadata'].get('policy')
-  assert feat_meta is not None
-  features_slice = feat_meta['output_slices']['hidden_state']
-  is_supercombo = vision_runner is None
+    derived_frame_skip = args.frame_skip or derive_frame_skip(vision_meta.get('input_shapes', {}), first_policy_meta.get('input_shapes', {}))
+    all_shapes = {key: value for meta in output_data['metadata'].values() for key, value in meta['input_shapes'].items()}
+    feat_meta = output_data['metadata'].get('vision') or output_data['metadata'].get('policy')
+    assert feat_meta is not None
+    features_slice = feat_meta['output_slices']['hidden_state']

-  print(f"Compiling run_policy JIT (model_size={model_w}x{model_h}, frame_skip={derived_frame_skip})...")
-  run_policy_func = make_run_policy(vision_runner, policy_runners, features_slice, derived_frame_skip, all_shapes)
-  run_policy_jit = TinyJit(run_policy_func, prune=True)
-  make_policy_queues = partial(generate_queues_and_npy, all_shapes, derived_frame_skip, is_supercombo=is_supercombo)
-  make_random_model_inputs = partial(make_random_images, keys=['warped'], shape=(2, 6, model_h // 2, model_w // 2), device=WARP_DEV)
-  output_data['run_policy'] = compile_jit(run_policy_jit, make_random_model_inputs, POLICY_INPUTS, make_policy_queues)
+    print(f"Compiling run_policy JIT (model_size={model_w}x{model_h}, frame_skip={derived_frame_skip})...")
+    run_policy_func = make_run_policy(vision_runner, policy_runners, features_slice, derived_frame_skip, all_shapes)
+    run_policy_jit = TinyJit(run_policy_func, prune=True)
+    make_policy_queues = partial(generate_queues_and_npy, all_shapes, derived_frame_skip, is_supercombo=False)
+    make_random_model_inputs = partial(make_random_images, keys=['warped'], shape=(2, 6, model_h // 2, model_w // 2), device=Device.DEFAULT)
+    output_data['run_policy'] = compile_jit(run_policy_jit, POLICY_INPUTS, make_policy_queues, make_random_inputs=make_random_model_inputs)

-  for cam_w, cam_h in args.camera_resolutions:
-    print(f"Compiling warp JIT for {cam_w}x{cam_h}...")
-    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 = TinyJit(make_warp(nv12, model_w, model_h), prune=True)
-    output_data[(cam_w, cam_h)] = compile_jit(warp, make_random_warp_inputs, WARP_INPUTS, make_warp_queues)
+    for cam_w, cam_h in args.camera_resolutions:
+      print(f"Compiling warp JIT for {cam_w}x{cam_h}...")
+      nv12 = stock.NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+      frame_copy_size = stock.nv12_copy_size(nv12.stride, nv12.y_height, nv12.uv_height)
+      make_random_warp_inputs = partial(make_random_images, keys=['frame', 'big_frame'], shape=frame_copy_size, device=Device.DEFAULT)
+      warp = TinyJit(stock.make_warp(nv12, model_w, model_h), prune=True)
+      output_data[(cam_w, cam_h)] = compile_jit(warp, WARP_INPUTS, make_warp_queues, make_random_inputs=make_random_warp_inputs)
+
+    output_data['metadata']['warp_dev'] = Device.DEFAULT

  with open(args.output, "wb") as file:
    dump_oob(output_data, file)
@@ -14,6 +14,8 @@ class ModelConstants:

  # model inputs constants
  MODEL_FREQ = 20
+  MODEL_RUN_FREQ = 20
+  MODEL_CONTEXT_FREQ = 5
  FEATURE_LEN = 512
  FULL_HISTORY_BUFFER_LEN = 99
  DESIRE_LEN = 8
@@ -35,6 +37,7 @@ 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
@@ -1,26 +1,9 @@
 from openpilot.sunnypilot.modeld_v2.constants import Meta
-from openpilot.cereal import custom
 from openpilot.sunnypilot.modeld_v2.meta_20hz import Meta20hz
 from openpilot.sunnypilot.models.helpers import get_active_bundle

-ModelBundle = custom.ModelManagerSP.ModelBundle
-

 def load_meta_constants():
-  """
-  Determines and loads the appropriate meta model class based on the metadata provided. The function checks
-  specific keys and conditions within the provided metadata dictionary to identify the corresponding meta
-  model class to return.
-
-  :param model_metadata: Dictionary containing metadata about the model. It includes
-      details such as input shapes, output slices, and other configurations for identifying
-      metadata-dependent meta model classes.
-  :type model_metadata: dict
-  :return: The appropriate meta model class (Meta, MetaSimPose, or MetaTombRaider)
-      based on the conditions and metadata provided.
-  :rtype: type
-  """
  if (bundle := get_active_bundle()) and bundle.is20hz:
    return Meta20hz
-
-  return Meta  # Default
+  return Meta
@@ -0,0 +1,187 @@
+"""
+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 argparse
+import os
+import sys
+import time
+import matplotlib.pyplot as plt
+import numpy as np
+from tinygrad.device import Device
+
+from openpilot.common.file_chunker import open_file_chunked
+from openpilot.selfdrive.modeld.compile_modeld import MODELD_INPUTS, make_input_queues, nv12_copy_size
+from openpilot.selfdrive.modeld.helpers import load_oob
+from openpilot.selfdrive.test.process_replay.model_replay import SEGMENT, TEST_ROUTE
+from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
+from openpilot.tools.lib.framereader import FrameReader
+from openpilot.tools.lib.openpilotci import get_url
+from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip
+from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser
+
+
+def get_replay_video_source(route_or_path=None, segment_index=SEGMENT, camera_type="fcamera.hevc"):
+  if route_or_path and os.path.exists(route_or_path):
+    return route_or_path
+  selected_route = route_or_path or TEST_ROUTE
+  return get_url(selected_route, segment_index, camera_type)
+
+
+def initialize_replay_queues(model_dictionary, device="CPU"):
+  metadata = model_dictionary.get("metadata", {})
+  model_meta = metadata.get("model", metadata)
+  input_shapes = model_meta.get("input_shapes", {})
+  cam_resolutions = list(model_dictionary.get("run_model", {}).keys())
+  cam_width, cam_height = cam_resolutions[0] if cam_resolutions else (1928, 1208)
+  nv12_info = get_nv12_info(cam_width, cam_height)
+  frame_copy_size = nv12_copy_size(nv12_info[0], nv12_info[1], nv12_info[2])
+  frame_skip = model_meta.get("frame_skip") or derive_frame_skip({}, input_shapes)
+  queues, npy_views, frame_views = make_input_queues(input_shapes, frame_skip, device, frame_copy_size)
+  if "tfm" in npy_views:
+    npy_views["tfm"][:] = np.eye(3, dtype=np.float32)
+  if "big_tfm" in npy_views:
+    npy_views["big_tfm"][:] = np.eye(3, dtype=np.float32)
+  if "traffic_convention" in npy_views:
+    npy_views["traffic_convention"][:] = np.array([1.0, 0.0], dtype=np.float32)
+  return queues, npy_views, frame_views, model_meta
+
+
+def replay_model_on_frames(model_path, frame_reader, number_of_frames=20):
+  with open_file_chunked(model_path) as file_handle:
+    model_data = load_oob(file_handle)
+  run_model_dict = model_data.get("run_model", {})
+  runner = next(iter(run_model_dict.values()), None)
+  if runner is None:
+    raise ValueError("Failed to resolve runner from model dictionary")
+
+  queues, npy_views, frame_views, model_meta = initialize_replay_queues(model_data)
+  output_slices = model_meta.get("output_slices", {})
+  hidden_state_slice = output_slices.get("hidden_state")
+  parser = Parser(ignore_missing=True)
+  recorded_outputs = []
+  max_frames = min(number_of_frames, getattr(frame_reader, "frame_count", number_of_frames))
+
+  for frame_index in range(max_frames):
+    frame_raw = frame_reader.get(frame_index)
+    if frame_raw is not None:
+      for view in frame_views.values():
+        copy_length = min(view.size, frame_raw.size)
+        view.flat[:copy_length] = frame_raw.flat[:copy_length]
+
+    execution_arguments = {key: queues[key] for key in MODELD_INPUTS if key in queues}
+    execution_start = time.perf_counter()
+    step_output = runner(**execution_arguments)
+    Device.default.synchronize()
+    step_duration = time.perf_counter() - execution_start
+
+    output_array = (step_output[0].numpy() if hasattr(step_output[0], "numpy") else np.array(step_output[0]))
+    flat_output = output_array.flatten()
+
+    if hidden_state_slice and "prev_feat" in npy_views:
+      features_flat = flat_output[hidden_state_slice]
+      target_slice = min(features_flat.size, npy_views["prev_feat"].size)
+      npy_views["prev_feat"].flat[:target_slice] = features_flat[:target_slice]
+
+    sliced_outputs = {slice_name: flat_output[np.newaxis, slice_range] for slice_name, slice_range in output_slices.items()}
+    parser.parse_outputs(sliced_outputs)
+
+    recorded_outputs.append({
+      "frame_index": frame_index,
+      "raw_output": output_array,
+      "parsed_outputs": sliced_outputs,
+      "execution_time": step_duration,
+    })
+  return recorded_outputs
+
+
+def plot_comparison(series_a, series_b, title, output_directory, label_a="modeld_v2 model", label_b="stock"):
+  os.makedirs(output_directory, exist_ok=True)
+  figure, axis = plt.subplots()
+  axis.plot(series_b, label=label_b)
+  axis.plot(series_a, label=label_a, linestyle="--")
+  axis.set_title(title)
+  axis.legend(loc="best")
+  plot_path = os.path.join(output_directory, f"{title}.png")
+  figure.savefig(plot_path)
+  plt.close(figure)
+  return plot_path
+
+
+def compare_models_on_route(new_model_path, old_model_path, route_or_path=None, segment_index=SEGMENT,
+                            number_of_frames=20, tolerance=1e-4, label_a="modeld_v2 model", label_b="stock",
+                            plot_directory=None, enforce_timings=False):
+  video_url_or_path = get_replay_video_source(route_or_path, segment_index)
+  frame_reader = FrameReader(video_url_or_path, pix_fmt="nv12")
+  old_results = replay_model_on_frames(old_model_path, frame_reader, number_of_frames)
+  new_results = replay_model_on_frames(new_model_path, frame_reader, number_of_frames)
+
+  for step_index, (new_step, old_step) in enumerate(zip(new_results, old_results, strict=True)):
+    new_array = new_step["raw_output"]
+    old_array = old_step["raw_output"]
+    if not np.allclose(new_array, old_array, atol=tolerance, rtol=tolerance):
+      max_absolute_error = np.max(np.abs(new_array - old_array))
+      sys.stderr.write(f"Replay mismatch at frame {step_index}: max absolute error {max_absolute_error:.6f} exceeds tolerance {tolerance}\n")
+      return False
+
+  if len(new_results) > 1 and len(old_results) > 1:
+    new_timings = [step["execution_time"] * 1000.0 for step in new_results[1:] if "execution_time" in step]
+    old_timings = [step["execution_time"] * 1000.0 for step in old_results[1:] if "execution_time" in step]
+    if new_timings and old_timings:
+      print("------------------------------------------------")
+      print("----------------- Model Timing -----------------")
+      print("------------------------------------------------")
+      print(f"{label_a}: avg {np.mean(new_timings):6.2f} ms | max {np.max(new_timings):6.2f} ms")
+      print(f"{label_b}: avg {np.mean(old_timings):6.2f} ms | max {np.max(old_timings):6.2f} ms")
+
+      if plot_directory:
+        first_step_outputs = new_results[0].get("parsed_outputs", {})
+        if "action" in first_step_outputs:
+          series_a_curv = [step["parsed_outputs"]["action"].flatten()[0] for step in new_results]
+          series_b_curv = [step["parsed_outputs"]["action"].flatten()[0] for step in old_results]
+          plot_comparison(series_a_curv, series_b_curv, "desiredCurvature", plot_directory, label_a, label_b)
+
+          series_a_accel = [step["parsed_outputs"]["action"].flatten()[1] for step in new_results]
+          series_b_accel = [step["parsed_outputs"]["action"].flatten()[1] for step in old_results]
+          plot_comparison(series_a_accel, series_b_accel, "desiredAcceleration", plot_directory, label_a, label_b)
+
+        if "plan" in first_step_outputs:
+          series_a_vel = [step["parsed_outputs"]["plan"].flatten()[0] for step in new_results]
+          series_b_vel = [step["parsed_outputs"]["plan"].flatten()[0] for step in old_results]
+          plot_comparison(series_a_vel, series_b_vel, "velocity.x", plot_directory, label_a, label_b)
+
+        if "lead" in first_step_outputs:
+          series_a_lead = [step["parsed_outputs"]["lead"].flatten()[0] for step in new_results]
+          series_b_lead = [step["parsed_outputs"]["lead"].flatten()[0] for step in old_results]
+          plot_comparison(series_a_lead, series_b_lead, "leadsV3.x", plot_directory, label_a, label_b)
+
+        plot_comparison(new_timings, old_timings, "execution_timings", plot_directory, label_a, label_b)
+
+        for slice_name in first_step_outputs:
+          series_a = [np.mean(step["parsed_outputs"][slice_name]) for step in new_results if slice_name in step["parsed_outputs"]]
+          series_b = [np.mean(step["parsed_outputs"][slice_name]) for step in old_results if slice_name in step["parsed_outputs"]]
+          if series_a and series_b:
+            plot_comparison(series_a, series_b, f"output_{slice_name}", plot_directory, label_a, label_b)
+  print(f"Replay comparison result on route ({label_a} vs {label_b}): True")
+  return True
+
+
+if __name__ == "__main__":
+  argument_parser = argparse.ArgumentParser(description="Model Replay on Real Driving Video")
+  argument_parser.add_argument("--sunnypilot-model", dest="model_a", default=None)
+  argument_parser.add_argument("--stock-model", dest="model_b", default=None)
+  argument_parser.add_argument("--route", default=TEST_ROUTE)
+  argument_parser.add_argument("--segment", type=int, default=SEGMENT)
+  argument_parser.add_argument("--frames", type=int, default=20)
+  argument_parser.add_argument("--plot-dir", default=None)
+  parsed_arguments = argument_parser.parse_args()
+
+  matches = compare_models_on_route(parsed_arguments.model_a, parsed_arguments.model_b, route_or_path=parsed_arguments.route,
+                                    segment_index=parsed_arguments.segment, number_of_frames=parsed_arguments.frames,
+                                    tolerance=1e-4, label_a="modeld_v2 model",
+                                    label_b="stock", plot_directory=parsed_arguments.plot_dir)
+  if not matches:
+    sys.exit(1)
@@ -6,6 +6,7 @@ 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.
 """

+from collections.abc import Callable
 import os
 os.environ['GMMU'] = '0'
 import numpy as np
@@ -37,11 +38,18 @@ from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value
 from openpilot.selfdrive.modeld.modeld import ChestnutState

+from openpilot.selfdrive.modeld.compile_modeld import (
+  MODELD_INPUTS,
+  make_input_queues as make_stock_input_queues,
+)
 from openpilot.sunnypilot.modeld_v2.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState, get_curvature_from_output
-from openpilot.sunnypilot.modeld_v2.constants import Plan
+from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser
+from openpilot.sunnypilot.modeld_v2.constants import ModelConstants, Plan
 from openpilot.sunnypilot.modeld_v2.meta_helper import load_meta_constants
 from openpilot.sunnypilot.modeld_v2.camera_offset_helper import CameraOffsetHelper
-from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues, make_supercombo_input_queues, WARP_INPUTS, POLICY_INPUTS
+from openpilot.sunnypilot.modeld_v2.compile_modeld import (derive_frame_skip, make_split_input_queues,
+                                                           make_supercombo_input_queues, nv12_copy_size,
+                                                           WARP_INPUTS, POLICY_INPUTS)
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
 from openpilot.sunnypilot.models.helpers import get_active_bundle
@@ -110,36 +118,40 @@ class ModelState(ModelStateBase):
    cloudlog.warning(f"loading combined pkl: {pkl_path}")
    jits = load_oob(open_file_chunked(pkl_path))

-    self.WARP_DEV = 'QCOM' if COMMA_HARDWARE else 'CPU'
-    self.DEV = 'AMD' if self.chestnut else self.WARP_DEV
-    self.QUEUE_DEV = self.DEV
    metadata = jits['metadata']
+    self.WARP_DEV = metadata.get('warp_dev', 'QCOM') if COMMA_HARDWARE else 'CPU'
+    self.DEV = ('AMD' if self.chestnut else 'QCOM') if COMMA_HARDWARE else 'CPU'
+    self.QUEUE_DEV = self.DEV
+    self.is_run_model = 'run_model' in jits

-    self.is_legacy_model = 'run_policy' not in jits  # remove after next recompile
-    if self.is_legacy_model:
-      self.warp = jits[(cam_w, cam_h)]['warp_enqueue']
-      self.run_policy = jits[(cam_w, cam_h)]['run_policy']
-    else:
-      self.run_policy = jits['run_policy']
-      self.warp = jits[(cam_w, cam_h)]
+    nv12_info = get_nv12_info(cam_w, cam_h)
+    self.frame_copy_size = nv12_copy_size(*nv12_info[:3])
+    self.full_frames: dict = {}
+    self._blob_cache: dict = {}
+    self.frame_buffers: dict = {}

-    if 'model' in metadata:
-      model_metadata = metadata['model']
+    if self.is_run_model or 'model' in metadata:
+      model_metadata = metadata.get('model', metadata)
+      self.input_shapes = model_metadata['input_shapes']
      self.vision_output_slices = model_metadata['output_slices']
      self.policy_output_slices = {}
      self._policy_slices_list = []
      self._combined_model_type = 'supercombo'
-      self._vision_input_names = [key for key in model_metadata['input_shapes'] if 'img' in key]
-      frame_skip = derive_frame_skip({}, model_metadata['input_shapes'])
-      self.input_queues, self.numpy_inputs = make_supercombo_input_queues(model_metadata['input_shapes'],
-                                                                          frame_skip, device=self.QUEUE_DEV)
-    else:
-      vision_metadata = metadata['vision']
-      policy_keys = [k for k in metadata if k != 'vision']
-      if policy_keys == ['policy']:
-        self._combined_model_type = 'split'
+      self._vision_input_names = [key for key in self.input_shapes if 'img' in key]
+      self.frame_skip = derive_frame_skip({}, self.input_shapes)
+      if self.is_run_model:
+        self.input_queues, self.numpy_inputs, self.frame_buffers = make_stock_input_queues(
+          self.input_shapes, self.frame_skip, device=self.DEV, frame_copy_size=self.frame_copy_size)
+        self.frame_views, self.npy = self.frame_buffers, self.numpy_inputs
+        self.run_model, self.run_policy, self.warp = jits['run_model'][(cam_w, cam_h)], None, None
      else:
-        self._combined_model_type = 'multi_policy'
+        self.input_queues, self.numpy_inputs = make_supercombo_input_queues(self.input_shapes, self.frame_skip, device=self.QUEUE_DEV)
+        self.run_model, self.run_policy, self.warp = None, jits['run_policy'], jits[(cam_w, cam_h)]
+    else:
+      self.run_model, self.run_policy, self.warp = None, jits['run_policy'], jits[(cam_w, cam_h)]
+      vision_metadata = metadata['vision']
+      policy_keys = [k for k in metadata if k not in ('vision', 'warp_dev')]
+      self._combined_model_type = 'split' if policy_keys == ['policy'] else 'multi_policy'
      self.vision_output_slices = vision_metadata['output_slices']
      self._policy_keys = policy_keys
      self._policy_slices_list = [metadata[k]['output_slices'] for k in policy_keys]
@@ -155,54 +167,39 @@ class ModelState(ModelStateBase):
    self._desire_key = next(key for key in self.numpy_inputs if key.startswith('desire'))
    self._road_key = next(key for key in self._vision_input_names if 'big' not in key)
    self._wide_key = next(key for key in self._vision_input_names if 'big' in key)
+    self.frame_buf_params = dict.fromkeys(self._vision_input_names, nv12_info)

    is_20hz = bundle.is20hz if bundle else self._combined_model_type in ('split', 'multi_policy')
    if is_20hz:
      from openpilot.sunnypilot.models.split_model_constants import SplitModelConstants
      self.constants = SplitModelConstants()
    else:
-      from openpilot.sunnypilot.modeld_v2.constants import ModelConstants
      self.constants = ModelConstants()

-    if self._combined_model_type != 'supercombo':
-      from openpilot.sunnypilot.modeld_v2.parse_model_outputs_split import Parser as SplitParser
-      self.parser = SplitParser()
-    else:
-      from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser as CombinedParser
-      self.parser = CombinedParser()
-
+    self.parser = Parser()
    self.prev_desire = np.zeros(self.constants.DESIRE_LEN, dtype=np.float32)
-    self.full_frames: dict = {}
-    self._blob_cache: dict = {}
-    nv12_info = get_nv12_info(cam_w, cam_h)
-    self.frame_buf_params = dict.fromkeys(self._vision_input_names, nv12_info)

-    yuv_size = self.frame_buf_params[self._road_key][3]
-    frame_tensor = Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize()
-    big_frame_tensor = Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize()
-
-    if self.is_legacy_model: # Remove this conditional hack after recompile
-      self.warp(**self.input_queues, frame=frame_tensor, big_frame=big_frame_tensor)
-    else:
-      self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=frame_tensor, big_frame=big_frame_tensor)
+    if self.warp is not None:
+      self.full_frames = {k: Tensor(np.zeros(nv12_info[3], dtype=np.uint8), device=self.WARP_DEV).contiguous().realize() for k in self._vision_input_names}
+      self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[self._road_key], big_frame=self.full_frames[self._wide_key])

  def warmup(self) -> None:
-    dummy_frames = {k: np.zeros(self.frame_buf_params[k][3], dtype=np.uint8) for k in self._vision_input_names}
+    dummy_size = self.frame_copy_size if self.is_run_model else self.frame_buf_params[self._road_key][3]
+    dummy_frames = {k: np.zeros(dummy_size, dtype=np.uint8) for k in self._vision_input_names}
    transforms = {k: np.eye(3, dtype=np.float32) for k in [self._road_key, self._wide_key] if k}
-
-    dummy_inputs = {}
-    for k, v in self.numpy_inputs.items():
-      if k not in ['tfm', 'big_tfm', 'prev_feat']:
-        dummy_inputs[k] = np.zeros(v.shape, dtype=v.dtype)
-
-    self.run(dummy_frames, transforms, dummy_inputs, prepare_only=False)
-
-    for v in self.numpy_inputs.values():
-      v[:] = 0
+    dummy_inputs = {k: np.zeros(v.shape, dtype=v.dtype) for k, v in self.numpy_inputs.items() if k not in ['tfm', 'big_tfm', 'prev_feat']}
+    self.run(dummy_frames, transforms, dummy_inputs)
+    if self.is_run_model:
+      self.input_queues, self.numpy_inputs, self.frame_buffers = make_stock_input_queues(
+        self.input_shapes, self.frame_skip, device=self.DEV, frame_copy_size=self.frame_copy_size)
+      self.frame_views = self.frame_buffers
+      self.npy = self.numpy_inputs
+    else:
+      for v in self.numpy_inputs.values():
+        v[:] = 0
+      self.full_frames.clear()
+      self._blob_cache.clear()
    self.prev_desire[:] = 0
-    self.full_frames.clear()
-    self._blob_cache.clear()
-

  @property
  def mlsim(self) -> bool:
@@ -217,45 +214,50 @@ class ModelState(ModelStateBase):
    return self._desire_key

  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:
-    for key in bufs.keys():
-      ptr = np.frombuffer(bufs[key].data, dtype=np.uint8).ctypes.data
-      yuv_size = self.frame_buf_params[key][3]
-      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', device=self.WARP_DEV)
-      self.full_frames[key] = self._blob_cache[cache_key]
+          inputs: dict[str, np.ndarray],
+          after_enqueue: Callable[[], None] | None = None) -> dict[str, np.ndarray] | None:
+    if self.is_run_model:
+      for key, buf in bufs.items():
+        data = buf.data if hasattr(buf, 'data') else buf
+        np.copyto(self.frame_buffers[key], np.frombuffer(data, dtype=np.uint8, count=self.frame_copy_size))
+    else:
+      for key, buf in bufs.items():
+        ptr = np.frombuffer(buf.data, dtype=np.uint8).ctypes.data
+        cache_key = (key, ptr)
+        if cache_key not in self._blob_cache:
+          self._blob_cache[cache_key] = Tensor.from_blob(ptr, (self.frame_buf_params[key][3],), dtype='uint8', device=self.WARP_DEV)
+        self.full_frames[key] = self._blob_cache[cache_key]

    desire_key = self.desire_key
    inputs[desire_key][0] = 0
    self.numpy_inputs[desire_key][:] = np.where(inputs[desire_key] - self.prev_desire > .99, inputs[desire_key], 0)
    self.prev_desire[:] = inputs[desire_key]
+
    for key in ('traffic_convention', 'lateral_control_params', 'action_t'):
      if key in self.numpy_inputs and key in inputs:
        self.numpy_inputs[key][:] = inputs[key]

-    road_key = self._road_key
-    wide_key = self._wide_key
-    self.numpy_inputs['tfm'][:, :] = transforms[road_key].reshape(3, 3)
-    self.numpy_inputs['big_tfm'][:, :] = transforms[wide_key].reshape(3, 3)
+    self.numpy_inputs['tfm'][:, :] = transforms[self._road_key].reshape(3, 3)
+    self.numpy_inputs['big_tfm'][:, :] = transforms[self._wide_key].reshape(3, 3)

-    if self.is_legacy_model:  # remove after next recompile
-      if prepare_only:
-        self.warp(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
-        return None
-      raw_outputs = self.run_policy(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+    if self.run_model is not None:
+      outs, = self.run_model(**{k: self.input_queues[k] for k in MODELD_INPUTS})
+      raw_outputs = outs
    else:
-      if prepare_only:
-        self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
-        return None
-      warped = self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+      assert self.warp is not None and self.run_policy is not None
+      warped = self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[self._road_key], big_frame=self.full_frames[self._wide_key])
      raw_outputs = self.run_policy(**{k: self.input_queues[k] for k in POLICY_INPUTS if k in self.input_queues}, warped=warped)

+    if after_enqueue is not None:
+      after_enqueue()
+
    if self._combined_model_type == 'supercombo':
      model_output = raw_outputs.numpy().flatten()
+      if self.chestnut and not np.all(np.isfinite(model_output)):
+        raise RuntimeError("model output not finite")
      sliced = {k: model_output[np.newaxis, v] for k, v in self.vision_output_slices.items()}
      outputs = self.parser.parse_outputs(sliced)
-      if 'prev_feat' in self.numpy_inputs:
+      if 'prev_feat' in self.numpy_inputs and 'hidden_state' in self.vision_output_slices:
        self.numpy_inputs['prev_feat'][:] = model_output[self.vision_output_slices['hidden_state']]
    else:
      vision_output = raw_outputs[0].numpy().flatten()
@@ -285,9 +287,6 @@ class ModelState(ModelStateBase):
      buf[0, :-1] = buf[0, 1:]
      buf[0, -1, :] = outputs['desired_curvature'][0, :] if not self.mlsim else 0

-    if self.chestnut and not np.all(np.isfinite(outputs.get('plan', np.array([0.])))):
-      raise RuntimeError("model output not finite")
-
    return outputs

  def get_action_from_model(self, model_output: dict[str, np.ndarray], prev_action: log.ModelDataV2.Action,
@@ -373,7 +372,11 @@ def main(demo=False):
    loader.start()
    loader.join(BIG_MODEL_TIMEOUT)
    model = big_model
+    if model is None:
+      params.put_bool("ChestnutModelError", True)
    params.put_bool("ChestnutActive", model is not None)
+    if model is not None:
+      params.remove("ChestnutModelError")

  small_model = ModelState(cam_w=vipc_client_main.width, cam_h=vipc_client_main.height, chestnut=False) if model is None or CHESTNUT else None
  if model is None:
@@ -487,9 +490,6 @@ def main(demo=False):
    run_count = run_count + 1

    frame_drop_ratio = frames_dropped / (1 + frames_dropped)
-    prepare_only = vipc_dropped_frames > 0
-    if prepare_only:
-      cloudlog.error(f"skipping model eval. Dropped {vipc_dropped_frames} frames")

    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}
@@ -512,11 +512,14 @@ def main(demo=False):

    mt1 = time.perf_counter()
    try:
-      model_output = model.run(bufs, transforms, inputs, prepare_only)
+      send_chestnut = (chestnut_state is not None and
+                       run_count % round(model.constants.MODEL_FREQ / SERVICE_LIST['chestnutState'].frequency) == 0)
+      model_output = model.run(bufs, transforms, inputs, chestnut_state.send if send_chestnut else None)
    except Exception:
      if not params.get_bool("ChestnutActive"):
        raise
      cloudlog.exception("chestnut failed, falling back to small")
+      params.put_bool("ChestnutModelError", True)
      params.put_bool("ChestnutActive", False)
      assert small_model is not None
      model = small_model
@@ -559,9 +562,6 @@ def main(demo=False):
      pm.send('modelDataV2SP', mdv2sp_send)
    last_vipc_frame_id = meta_main.frame_id

-    if chestnut_state is not None and run_count % round(model.constants.MODEL_FREQ / SERVICE_LIST['chestnutState'].frequency) == 0:
-      chestnut_state.send()
-
 if __name__ == "__main__":
  try:
    import argparse
@@ -115,22 +115,41 @@ class Parser:
    outs[name + '_stds'] = pred_std_final.reshape(final_shape)

  def parse_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
-    # supercombo (4955 / 102) and newer variants (e.g. 990 / 144).
-    self.parse_mdn('plan',   outs, out_shape=(ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH))
-    self.parse_mdn('lane_lines', outs, out_shape=(ModelConstants.NUM_LANE_LINES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
-    self.parse_mdn('road_edges', outs, out_shape=(ModelConstants.NUM_ROAD_EDGES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
-    self.parse_mdn('pose',         outs, out_shape=(ModelConstants.POSE_WIDTH,))
-    self.parse_mdn('road_transform', outs, out_shape=(ModelConstants.POSE_WIDTH,))
+    if 'plan' in outs:
+      self.parse_mdn('plan', outs, out_shape=(ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH))
+    if 'planplus' in outs:
+      self.parse_mdn('planplus', outs, out_shape=(ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH))
+    if 'lane_lines' in outs:
+      self.parse_mdn('lane_lines', outs, out_shape=(ModelConstants.NUM_LANE_LINES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
+    if 'road_edges' in outs:
+      self.parse_mdn('road_edges', outs, out_shape=(ModelConstants.NUM_ROAD_EDGES, ModelConstants.IDX_N, ModelConstants.LANE_LINES_WIDTH))
+    if 'pose' in outs:
+      self.parse_mdn('pose', outs, out_shape=(ModelConstants.POSE_WIDTH,))
+    if 'road_transform' in outs:
+      self.parse_mdn('road_transform', outs, out_shape=(ModelConstants.POSE_WIDTH,))
    if 'sim_pose' in outs:
      self.parse_mdn('sim_pose', outs, out_shape=(ModelConstants.POSE_WIDTH,))
-    self.parse_mdn('wide_from_device_euler', outs, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
-    self.parse_mdn('lead', outs, out_shape=(ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH))
+    if 'wide_from_device_euler' in outs:
+      self.parse_mdn('wide_from_device_euler', outs, out_shape=(ModelConstants.WIDE_FROM_DEVICE_WIDTH,))
+    if 'lead' in outs:
+      self.parse_mdn('lead', outs, out_shape=(ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH))
    if 'lat_planner_solution' in outs:
      self.parse_mdn('lat_planner_solution', outs, out_shape=(ModelConstants.IDX_N, ModelConstants.LAT_PLANNER_SOLUTION_WIDTH))
    if 'desired_curvature' in outs:
      self.parse_mdn('desired_curvature', outs, out_shape=(ModelConstants.DESIRED_CURV_WIDTH,))
+    if 'action' in outs:
+      self.parse_mdn('action', outs, out_shape=(ModelConstants.ACTION_WIDTH,))
    for k in ['lead_prob', 'lane_lines_prob', 'meta']:
-      self.parse_binary_crossentropy(k, outs)
-    self.parse_categorical_crossentropy('desire_state', outs, out_shape=(ModelConstants.DESIRE_PRED_WIDTH,))
-    self.parse_categorical_crossentropy('desire_pred',  outs, out_shape=(ModelConstants.DESIRE_PRED_LEN, ModelConstants.DESIRE_PRED_WIDTH))
+      if k in outs:
+        self.parse_binary_crossentropy(k, outs)
+    if 'desire_state' in outs:
+      self.parse_categorical_crossentropy('desire_state', outs, out_shape=(ModelConstants.DESIRE_PRED_WIDTH,))
+    if 'desire_pred' in outs:
+      self.parse_categorical_crossentropy('desire_pred', outs, out_shape=(ModelConstants.DESIRE_PRED_LEN, ModelConstants.DESIRE_PRED_WIDTH))
    return outs
+
+  def parse_vision_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    return self.parse_outputs(outs)
+
+  def parse_policy_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+    return self.parse_outputs(outs)
@@ -1,159 +0,0 @@
-import numpy as np
-from openpilot.sunnypilot.models.split_model_constants import SplitModelConstants
-
-
-def safe_exp(x, out=None):
-  # -11 is around 10**14, more causes float16 overflow
-  return np.exp(np.clip(x, -np.inf, 11), out=out)
-
-
-def sigmoid(x):
-  return 1. / (1. + safe_exp(-x))
-
-
-def softmax(x, axis=-1):
-  x -= np.max(x, axis=axis, keepdims=True)
-  if x.dtype == np.float32 or x.dtype == np.float64:
-    safe_exp(x, out=x)
-  else:
-    x = safe_exp(x)
-  x /= np.sum(x, axis=axis, keepdims=True)
-  return x
-
-
-class Parser:
-  def __init__(self, ignore_missing=False):
-    self.ignore_missing = ignore_missing
-
-  def check_missing(self, outs, name):
-    if name not in outs and not self.ignore_missing:
-      raise ValueError(f"Missing output {name}")
-    return name not in outs
-
-  def parse_categorical_crossentropy(self, name, outs, out_shape=None):
-    if self.check_missing(outs, name):
-      return
-    raw = outs[name]
-    if out_shape is not None:
-      raw = raw.reshape((raw.shape[0],) + out_shape)
-    outs[name] = softmax(raw, axis=-1)
-
-  def parse_binary_crossentropy(self, name, outs):
-    if self.check_missing(outs, name):
-      return
-    raw = outs[name]
-    outs[name] = sigmoid(raw)
-
-  def parse_mdn(self, name, outs, in_N=0, out_N=1, out_shape=None):
-    if self.check_missing(outs, name):
-      return
-    raw = outs[name]
-    raw = raw.reshape((raw.shape[0], max(in_N, 1), -1))
-
-    n_values = (raw.shape[2] - out_N)//2
-    pred_mu = raw[:,:,:n_values]
-    pred_std = safe_exp(raw[:,:,n_values: 2*n_values])
-
-    if in_N > 1:
-      weights = np.zeros((raw.shape[0], in_N, out_N), dtype=raw.dtype)
-      for i in range(out_N):
-        weights[:,:,i - out_N] = softmax(raw[:,:,i - out_N], axis=-1)
-
-      if out_N == 1:
-        for fidx in range(weights.shape[0]):
-          idxs = np.argsort(weights[fidx][:,0])[::-1]
-          weights[fidx] = weights[fidx][idxs]
-          pred_mu[fidx] = pred_mu[fidx][idxs]
-          pred_std[fidx] = pred_std[fidx][idxs]
-      assert out_shape is not None
-      full_shape = tuple([raw.shape[0], in_N] + list(out_shape))
-      outs[name + '_weights'] = weights
-      outs[name + '_hypotheses'] = pred_mu.reshape(full_shape)
-      outs[name + '_stds_hypotheses'] = pred_std.reshape(full_shape)
-
-      pred_mu_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
-      pred_std_final = np.zeros((raw.shape[0], out_N, n_values), dtype=raw.dtype)
-      for fidx in range(weights.shape[0]):
-        for hidx in range(out_N):
-          idxs = np.argsort(weights[fidx,:,hidx])[::-1]
-          pred_mu_final[fidx, hidx] = pred_mu[fidx, idxs[0]]
-          pred_std_final[fidx, hidx] = pred_std[fidx, idxs[0]]
-    else:
-      pred_mu_final = pred_mu
-      pred_std_final = pred_std
-
-    if out_N > 1:
-      assert out_shape is not None
-      final_shape = tuple([raw.shape[0], out_N] + list(out_shape))
-    else:
-      assert out_shape is not None
-      final_shape = tuple([raw.shape[0],] + list(out_shape))
-    outs[name] = pred_mu_final.reshape(final_shape)
-    outs[name + '_stds'] = pred_std_final.reshape(final_shape)
-
-  def is_mhp(self, outs, name, shape):
-    if self.check_missing(outs, name):
-      return False
-    if outs[name].shape[1] == 2 * shape:
-      return False
-    return True
-
-  def parse_dynamic_outputs(self, outs: dict[str, np.ndarray]) -> None:
-    if 'lead' in outs:
-      lead_mhp = self.is_mhp(outs, 'lead',
-                             SplitModelConstants.LEAD_MHP_SELECTION * SplitModelConstants.LEAD_TRAJ_LEN * SplitModelConstants.LEAD_WIDTH)
-      lead_in_N, lead_out_N = (SplitModelConstants.LEAD_MHP_N, SplitModelConstants.LEAD_MHP_SELECTION) if lead_mhp else (0, 0)
-      lead_out_shape = (SplitModelConstants.LEAD_TRAJ_LEN, SplitModelConstants.LEAD_WIDTH) if lead_mhp else \
-        (SplitModelConstants.LEAD_MHP_SELECTION, SplitModelConstants.LEAD_TRAJ_LEN, SplitModelConstants.LEAD_WIDTH)
-      self.parse_mdn('lead', outs, in_N=lead_in_N, out_N=lead_out_N, out_shape=lead_out_shape)
-    if 'plan' in outs:
-      plan_mhp = self.is_mhp(outs, 'plan', SplitModelConstants.IDX_N * SplitModelConstants.PLAN_WIDTH)
-      plan_in_N, plan_out_N = (SplitModelConstants.PLAN_MHP_N, SplitModelConstants.PLAN_MHP_SELECTION) if plan_mhp else (0, 0)
-      self.parse_mdn('plan', outs, in_N=plan_in_N, out_N=plan_out_N,
-                     out_shape=(SplitModelConstants.IDX_N, SplitModelConstants.PLAN_WIDTH))
-    if 'planplus' in outs:
-      self.parse_mdn('planplus', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.IDX_N, SplitModelConstants.PLAN_WIDTH))
-
-  def split_outputs(self, outs: dict[str, np.ndarray]) -> None:
-    if 'desired_curvature' in outs:
-      self.parse_mdn('desired_curvature', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.DESIRED_CURV_WIDTH,))
-    if 'desire_pred' in outs:
-      self.parse_categorical_crossentropy('desire_pred', outs, out_shape=(SplitModelConstants.DESIRE_PRED_LEN,SplitModelConstants.DESIRE_PRED_WIDTH))
-    if 'desire_state' in outs:
-      self.parse_categorical_crossentropy('desire_state', outs, out_shape=(SplitModelConstants.DESIRE_PRED_WIDTH,))
-    if 'lane_lines' in outs:
-      self.parse_mdn('lane_lines', outs, in_N=0, out_N=0,
-                     out_shape=(SplitModelConstants.NUM_LANE_LINES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
-    if 'lane_lines_prob' in outs:
-      self.parse_binary_crossentropy('lane_lines_prob', outs)
-    if 'lead_prob' in outs:
-      self.parse_binary_crossentropy('lead_prob', outs)
-    if 'lat_planner_solution' in outs:
-      self.parse_mdn('lat_planner_solution', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.IDX_N,SplitModelConstants.LAT_PLANNER_SOLUTION_WIDTH))
-    if 'meta' in outs:
-      self.parse_binary_crossentropy('meta', outs)
-    if 'road_edges' in outs:
-      self.parse_mdn('road_edges', outs, in_N=0, out_N=0,
-                     out_shape=(SplitModelConstants.NUM_ROAD_EDGES,SplitModelConstants.IDX_N,SplitModelConstants.LANE_LINES_WIDTH))
-    if 'sim_pose' in outs:
-      self.parse_mdn('sim_pose', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.POSE_WIDTH,))
-    if 'action' in outs:
-      self.parse_mdn('action', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.ACTION_WIDTH,))
-
-  def parse_vision_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
-    self.parse_mdn('pose', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.POSE_WIDTH,))
-    self.parse_mdn('wide_from_device_euler', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.WIDE_FROM_DEVICE_WIDTH,))
-    self.parse_mdn('road_transform', outs, in_N=0, out_N=0, out_shape=(SplitModelConstants.POSE_WIDTH,))
-    self.parse_dynamic_outputs(outs)
-    self.split_outputs(outs)
-    return outs
-
-  def parse_policy_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
-    self.parse_dynamic_outputs(outs)
-    self.split_outputs(outs)
-    return outs
-
-  def parse_outputs(self, outs: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
-    outs = self.parse_vision_outputs(outs)
-    outs = self.parse_policy_outputs(outs)
-    return outs
@@ -117,7 +117,7 @@ ARCHETYPES = {
    is_20hz=True,
    expected_model_type='split',
    expected_constants_class=SplitModelConstants,
-    expected_parser_module='parse_model_outputs_split',
+    expected_parser_module='parse_model_outputs',
    expected_desire_key='desire',
  ),
  'vision_multi_policy': Archetype(
@@ -130,7 +130,7 @@ ARCHETYPES = {
    is_20hz=True,
    expected_model_type='multi_policy',
    expected_constants_class=SplitModelConstants,
-    expected_parser_module='parse_model_outputs_split',
+    expected_parser_module='parse_model_outputs',
    expected_desire_key='desire',
  ),
  'tri_policy': Archetype(
@@ -144,7 +144,7 @@ ARCHETYPES = {
    is_20hz=True,
    expected_model_type='multi_policy',
    expected_constants_class=SplitModelConstants,
-    expected_parser_module='parse_model_outputs_split',
+    expected_parser_module='parse_model_outputs',
    expected_desire_key='desire',
  ),
  'supercombo_non20hz': Archetype(
@@ -103,6 +103,23 @@ class TestStockEquivalence(OpenpilotTestCase):
    assert state.vision_output_slices == arch.metadata_structure['vision']['output_slices']
    assert state.policy_output_slices == arch.metadata_structure['policy']['output_slices']

+  def test_unified_run_model(self, tmp_path, monkeypatch, patch_modeld):
+    from openpilot.common.hardware import hw
+    from openpilot.selfdrive.modeld.helpers import dump_oob
+    shapes = {'img': (1, 12, 128, 256), 'big_img': (1, 12, 128, 256), 'features_buffer': (1, 24, 32, 512),
+              'desire_pulse': (1, 25, 8), 'traffic_convention': (1, 2), 'action_t': (1, 2)}
+    pkl_data = {'metadata': {'model': {'input_shapes': shapes, 'output_slices': {}}},
+                'run_model': {(CAM_W, CAM_H): tests_helpers._noop_jit}}
+    with open(tmp_path / 'driving_test_tinygrad.pkl', 'wb') as f:
+      dump_oob(pkl_data, f)
+    bundle = DummyBundle(models=[DummyModel('supercombo', 'driving_test_tinygrad.pkl')])
+    patch_modeld(bundle)
+    monkeypatch.setattr(hw.Paths, 'model_root', staticmethod(lambda: str(tmp_path)))
+    state = ModelState(cam_w=CAM_W, cam_h=CAM_H)
+    assert state.is_run_model and state.run_model is not None
+    assert state.run_policy is None and state.warp is None
+    assert 'img' in state.frame_views and 'big_img' in state.frame_views
+

 ARCHETYPE_NAMES = list(ARCHETYPES.keys())

@@ -0,0 +1,81 @@
+import numpy as np
+from openpilot.common.test import OpenpilotTestCase
+from openpilot.sunnypilot.modeld_v2.constants import ModelConstants
+from openpilot.sunnypilot.modeld_v2.parse_model_outputs import Parser, _infer_mhp, sigmoid, softmax
+
+
+class TestParseModelOutputs(OpenpilotTestCase):
+  def test_infer_mhp_lead(self):
+    in_hypotheses, out_selections = _infer_mhp(102, 24)
+    assert in_hypotheses == 2
+    assert out_selections == 3
+
+  def test_infer_mhp_plan(self):
+    in_hypotheses, out_selections = _infer_mhp(4955, 495)
+    assert in_hypotheses == 5
+    assert out_selections == 1
+
+  def test_infer_mhp_non_mdn(self):
+    in_hypotheses, out_selections = _infer_mhp(48, 24)
+    assert in_hypotheses == 1
+    assert out_selections == 0
+
+  def test_check_missing_raises(self):
+    parser = Parser(ignore_missing=False)
+    with self.assertRaises(ValueError):
+      parser.check_missing({}, "missing_key")
+
+  def test_check_missing_ignored(self):
+    parser = Parser(ignore_missing=True)
+    assert parser.check_missing({}, "missing_key") is True
+
+  def test_binary_crossentropy(self):
+    parser = Parser()
+    raw_logits = np.array([[-10.0, 0.0, 10.0]], dtype=np.float32)
+    outs = {"meta": raw_logits.copy()}
+    parser.parse_binary_crossentropy("meta", outs)
+    expected_probabilities = sigmoid(raw_logits)
+    np.testing.assert_allclose(outs["meta"], expected_probabilities, rtol=1e-5, atol=1e-6)
+
+  def test_categorical_crossentropy(self):
+    parser = Parser()
+    raw_logits = np.array([[1.0, 2.0, 3.0]], dtype=np.float32)
+    outs = {"desire_state": raw_logits.copy()}
+    parser.parse_categorical_crossentropy("desire_state", outs)
+    expected_probabilities = softmax(raw_logits)
+    np.testing.assert_allclose(outs["desire_state"], expected_probabilities, rtol=1e-5, atol=1e-6)
+
+  def test_parse_vision_outputs(self):
+    parser = Parser()
+    pose_raw = np.zeros((1, ModelConstants.POSE_WIDTH * 2), dtype=np.float32)
+    road_transform_raw = np.zeros((1, ModelConstants.POSE_WIDTH * 2), dtype=np.float32)
+    lead_raw = np.zeros((1, 102), dtype=np.float32)
+    meta_raw = np.zeros((1, 55), dtype=np.float32)
+    vision_outputs = {"pose": pose_raw, "road_transform": road_transform_raw, "lead": lead_raw, "meta": meta_raw}
+    parsed = parser.parse_vision_outputs(vision_outputs)
+    assert "pose" in parsed
+    assert "road_transform" in parsed
+    assert "lead" in parsed
+    assert "meta" in parsed
+    assert parsed["pose"].shape == (1, ModelConstants.POSE_WIDTH)
+    assert parsed["lead"].shape == (1, ModelConstants.LEAD_MHP_SELECTION, ModelConstants.LEAD_TRAJ_LEN, ModelConstants.LEAD_WIDTH)
+
+  def test_parse_policy_outputs(self):
+    parser = Parser()
+    plan_raw = np.zeros((1, 4955), dtype=np.float32)
+    desire_state_raw = np.zeros((1, ModelConstants.DESIRE_PRED_WIDTH), dtype=np.float32)
+    action_raw = np.zeros((1, ModelConstants.ACTION_WIDTH * 2), dtype=np.float32)
+    policy_outputs = {"plan": plan_raw, "desire_state": desire_state_raw, "action": action_raw}
+    parsed = parser.parse_policy_outputs(policy_outputs)
+    assert parsed["plan"].shape == (1, ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH)
+    assert parsed["action"].shape == (1, ModelConstants.ACTION_WIDTH)
+    assert parsed["desire_state"].shape == (1, ModelConstants.DESIRE_PRED_WIDTH)
+
+  def test_parse_outputs_combined(self):
+    parser = Parser()
+    outputs = {"plan": np.zeros((1, 4955), dtype=np.float32), "pose": np.zeros((1, ModelConstants.POSE_WIDTH * 2),
+               dtype=np.float32), "meta": np.zeros((1, 55), dtype=np.float32)}
+    parsed = parser.parse_outputs(outputs)
+    assert parsed["plan"].shape == (1, ModelConstants.IDX_N, ModelConstants.PLAN_WIDTH)
+    assert parsed["pose"].shape == (1, ModelConstants.POSE_WIDTH)
+    assert parsed["meta"].shape == (1, 55)
@@ -1,121 +0,0 @@
-import numpy as np
-
-def index_function(idx, max_val=192, max_idx=32):
-  return max_val * ((idx/max_idx)**2)
-
-
-class ModelConstants:
-  # time and distance indices
-  IDX_N = 33
-  T_IDXS = [index_function(idx, max_val=10.0) for idx in range(IDX_N)]
-  X_IDXS = [index_function(idx, max_val=192.0) for idx in range(IDX_N)]
-  LEAD_T_IDXS = [0., 2., 4., 6., 8., 10.]
-  LEAD_T_OFFSETS = [0., 2., 4.]
-  META_T_IDXS = [2., 4., 6., 8., 10.]
-
-  # model inputs constants
-  MODEL_FREQ = 20
-  FEATURE_LEN = 512
-  HISTORY_BUFFER_LEN = 99
-  DESIRE_LEN = 8
-  TRAFFIC_CONVENTION_LEN = 2
-  NAV_FEATURE_LEN = 256
-  NAV_INSTRUCTION_LEN = 150
-  LAT_PLANNER_STATE_LEN = 4
-  LATERAL_CONTROL_PARAMS_LEN = 2
-  PREV_DESIRED_CURV_LEN = 1
-
-  # model outputs constants
-  FCW_THRESHOLDS_5MS2 = np.array([.05, .05, .15, .15, .15], dtype=np.float32)
-  FCW_THRESHOLDS_3MS2 = np.array([.7, .7], dtype=np.float32)
-  FCW_5MS2_PROBS_WIDTH = 5
-  FCW_3MS2_PROBS_WIDTH = 2
-
-  DISENGAGE_WIDTH = 5
-  POSE_WIDTH = 6
-  WIDE_FROM_DEVICE_WIDTH = 3
-  SIM_POSE_WIDTH = 6
-  LEAD_WIDTH = 4
-  LANE_LINES_WIDTH = 2
-  ROAD_EDGES_WIDTH = 2
-  PLAN_WIDTH = 15
-  DESIRE_PRED_WIDTH = 8
-  LAT_PLANNER_SOLUTION_WIDTH = 4
-  DESIRED_CURV_WIDTH = 1
-
-  NUM_LANE_LINES = 4
-  NUM_ROAD_EDGES = 2
-
-  LEAD_TRAJ_LEN = 6
-  DESIRE_PRED_LEN = 4
-
-  PLAN_MHP_N = 5
-  LEAD_MHP_N = 2
-  PLAN_MHP_SELECTION = 1
-  LEAD_MHP_SELECTION = 3
-
-  FCW_THRESHOLD_5MS2_HIGH = 0.15
-  FCW_THRESHOLD_5MS2_LOW = 0.05
-  FCW_THRESHOLD_3MS2 = 0.7
-
-  CONFIDENCE_BUFFER_LEN = 5
-  RYG_GREEN = 0.01165
-  RYG_YELLOW = 0.06157
-
-  POLY_PATH_DEGREE = 4
-
-
-# model outputs slices
-class Plan:
-  POSITION = slice(0, 3)
-  VELOCITY = slice(3, 6)
-  ACCELERATION = slice(6, 9)
-  T_FROM_CURRENT_EULER = slice(9, 12)
-  ORIENTATION_RATE = slice(12, 15)
-
-
-class Meta:
-  ENGAGED = slice(0, 1)
-  # next 2, 4, 6, 8, 10 seconds
-  GAS_DISENGAGE = slice(1, 31, 6)
-  BRAKE_DISENGAGE = slice(2, 31, 6)
-  STEER_OVERRIDE = slice(3, 31, 6)
-  HARD_BRAKE_3 = slice(4, 31, 6)
-  HARD_BRAKE_4 = slice(5, 31, 6)
-  HARD_BRAKE_5 = slice(6, 31, 6)
-  # next 0, 2, 4, 6, 8, 10 seconds
-  GAS_PRESS = slice(31, 55, 4)
-  BRAKE_PRESS = slice(32, 55, 4)
-  LEFT_BLINKER = slice(33, 55, 4)
-  RIGHT_BLINKER = slice(34, 55, 4)
-
-
-class MetaTombRaider:
-  ENGAGED = slice(0, 1)
-  # next 2, 4, 6, 8, 10 seconds
-  GAS_DISENGAGE = slice(1, 41, 8)
-  BRAKE_DISENGAGE = slice(2, 41, 8)
-  STEER_OVERRIDE = slice(3, 41, 8)
-  HARD_BRAKE_3 = slice(4, 41, 8)
-  HARD_BRAKE_4 = slice(5, 41, 8)
-  HARD_BRAKE_5 = slice(6, 41, 8)
-  GAS_PRESS = slice(7, 41, 8)
-  BRAKE_PRESS = slice(8, 41, 8)
-  # next 0, 2, 4, 6, 8, 10 seconds
-  LEFT_BLINKER = slice(41, 53, 2)
-  RIGHT_BLINKER = slice(42, 53, 2)
-
-
-class MetaSimPose:
-  ENGAGED = slice(0, 1)
-  # next 2, 4, 6, 8, 10 seconds
-  GAS_DISENGAGE = slice(1, 36, 7)
-  BRAKE_DISENGAGE = slice(2, 36, 7)
-  STEER_OVERRIDE = slice(3, 36, 7)
-  HARD_BRAKE_3 = slice(4, 36, 7)
-  HARD_BRAKE_4 = slice(5, 36, 7)
-  HARD_BRAKE_5 = slice(6, 36, 7)
-  GAS_PRESS = slice(7, 36, 7)
-  # next 0, 2, 4, 6, 8, 10 seconds
-  LEFT_BLINKER = slice(36, 48, 2)
-  RIGHT_BLINKER = slice(37, 48, 2)
@@ -139,7 +139,7 @@ class ModelCache:
 class ModelFetcher:
  """Handles fetching and caching of model data from remote source"""
  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v22.json"
-  MODEL_URL_CHESTNUT = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_chestnut_v23.json"
+  MODEL_URL_CHESTNUT = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_chestnut_v25.json"

  MODEL_SOURCES = {
    "qcom": (MODEL_URL, ""),
@@ -7,14 +7,11 @@ See the LICENSE.md file in the root directory for more details.

 import hashlib
 import os
-import pickle
-from pathlib import Path
 import numpy as np

 from openpilot.cereal import custom
 from openpilot.common.params import Params
 from openpilot.common.swaglog import cloudlog
-from openpilot.sunnypilot.models.constants import Meta, MetaSimPose, MetaTombRaider
 from openpilot.common.hardware.hw import Paths
 from openpilot.selfdrive.modeld.helpers import chestnut_present

@@ -22,7 +19,6 @@ from openpilot.selfdrive.modeld.helpers import chestnut_present
 REQUIRED_JSON_VERSION = 19

 CUSTOM_MODEL_PATH = Paths.model_root()
-METADATA_PATH = Path(__file__).parent / '../models/supercombo_metadata.pkl'
 ModelManager = custom.ModelManagerSP

 ACTIVE_BUNDLE_KEYS = {
@@ -201,33 +197,6 @@ def _get_model():
  return None


-def load_metadata():
-  metadata_path = METADATA_PATH
-
-  with open(metadata_path, 'rb') as f:
-    return pickle.load(f)
-
-
-def prepare_inputs(model_metadata: dict) -> dict[str, np.ndarray]:
-  return {
-    key: np.zeros(shape, dtype=np.float32).flatten()
-    for key, shape in model_metadata['input_shapes'].items()
-    if 'img' not in key
-  }
-
-
-def load_meta_constants(model_metadata: dict):
-  """ Loads the appropriate meta model class based on key shapes"""
-  if 'sim_pose' in model_metadata['input_shapes']:
-    return MetaSimPose
-
-  meta_slice = model_metadata['output_slices']['meta']
-  if (meta_slice.start, meta_slice.stop, meta_slice.step) == (5868, 5921, None):
-    return MetaTombRaider
-
-  return Meta
-
-
 # The following method(s) are modeld helper methods
 def plan_x_idxs_helper(constants, plan, model_output) -> list[float]:
  # times at X_IDXS according to plan.
@@ -38,7 +38,8 @@ def main():
  api = HfApi()
  onnx_sha256 = hash_file(args.onnx_path)
  short_ref = args.onnx_ref[:8]
-  folder_name = f"model-{args.model_name}-{short_ref}-{args.run_number}"
+  safe_name = args.model_name.replace(" ", "-")
+  folder_name = f"model-{safe_name}-{short_ref}-{args.run_number}"

  print(f"ONNX hash: {onnx_sha256}")
  print(f"ONNX ref: {args.onnx_ref} (short: {short_ref})")
Author	SHA1	Message	Date
discountchubbs	429a334d60	clean up some MORE	2026-09-09 08:58:19 -07:00
James Vecellio-Grant	c0a3ca44b6	Update model_replay.py	2026-09-09 00:56:07 -07:00
discountchubbs	a039d9bc27	rm	2026-09-09 00:24:09 -07:00
discountchubbs	c634f6ed8e	y no more nproc	2026-09-09 00:08:41 -07:00
discountchubbs	fd075878dd	eh use latest. it has m1 chip	2026-09-09 00:05:23 -07:00
discountchubbs	5a62f7b427	drop	2026-09-09 00:01:03 -07:00
discountchubbs	13ea924578	wrap both models	2026-09-08 23:39:36 -07:00
discountchubbs	ab07b706a2	lil more	2026-09-08 23:28:40 -07:00
discountchubbs	50ea3a0388	clean	2026-09-08 23:25:29 -07:00
discountchubbs	1371618685	Update model_replay.yaml	2026-09-08 23:17:20 -07:00
discountchubbs	3f3e918a42	compile both at same time FULL SPEED AHEAD	2026-09-08 23:09:14 -07:00
discountchubbs	c625b719c1	Update model_replay.yaml	2026-09-08 23:01:21 -07:00
discountchubbs	86f67d7aa9	Update model_replay.yaml	2026-09-08 22:59:21 -07:00
discountchubbs	857eb5e135	Update model_replay.yaml	2026-09-08 22:47:42 -07:00
discountchubbs	c2d0b415af	replay deez 🌰	2026-09-08 22:42:28 -07:00
James Vecellio-Grant	6135084c94	modeld_v2: realize frames on npy -> amd (#1993 )	2026-09-05 20:17:14 -07:00
James Vecellio-Grant	047ae41c0d	modeld_v2: one dev warp and enqueue (#1990 )	2026-09-04 21:15:00 -07:00
Nayan	7eb457f6c4	chaos (#1991 ) burn it all	2026-09-05 10:50:12 +08:00
James Vecellio-Grant	302f3ad892	ci: compile dm warp (#1989 )	2026-09-04 16:52:09 -07:00
Jason Wen	132b31f4cf	ci: poll GH API in prepare model jobs (#1987 )	2026-09-03 10:10:21 -04:00
James Vecellio-Grant	752c07f9e4	ci: Replace hf oath with token (#1986 ) replace oauth with token	2026-09-03 08:22:43 -04:00
Jason Wen	e87dbbaba7	models: sanitize default model name for HF (#1984 )	2026-09-02 14:53:31 -04:00
Jason Wen	15efdb392f	Sync: `commaai/openpilot:master` → `sunnypilot/sunnypilot:master` (#1983 ) * ui: remove raygui usage (#38708) * ui: remove raygui usage * match previous gui_text_box line spacing * Revert "match previous gui_text_box line spacing" This reverts commit ffd2fe31725c6d50bffaebc621c1e170d0926c66. * Reapply "match previous gui_text_box line spacing" This reverts commit d41404f09607e225f43868f7747f22dc0bb2cf16. * log chestnut supply fault (#38711) * log chestnut INA supply fault * ci * bump raylib (#38712) * cabana: replace custom non-view Qt signals w/ plain observer (#38713) * cabana: move RoutesDialog out of streams/ (#38716) * cabana: string helpers in utils return std::string (#38720) * cabana: use std::string in RoutesDialog API results (#38717) * cabana: move stream open widgets into streamselector (#38715) * cabana: remove Qt from livestream (#38722) * cabana: split SettingsDialog out of settings (#38719) cabana: split SettingsDialog out of settings.{h,cc} * cabana: split comma API route fetching out of RoutesDialog (#38721) * cabana: de-QT streams (#38718) * ui: fix install update button overflow (#38696) * cabana: split utils/util into Qt-free util and qtutil (#38723) * ui: guard branch switcher before internet connected (#38692) * ui: check for update on target branch switch (#38693) * ui: sync gpu loading to offroad (#38727) ui: sync gpu loading state * add chestnut offroad alerts (#38706) * system: add chestnut offroad alerts * system: refine chestnut offroad alerts * system: refine chestnut power alerts * system: confirm chestnut power recovery from PCIe * system: detect missing chestnut power from INA voltage * common: fix OpenpilotPrefix cleanup on macOS (#38728) The destructor built its cleanup commands as "rm %s -rf", with the flags after the operand. GNU rm permutes arguments so this works on device and in CI, but BSD rm on macOS stops option parsing at the first operand and treats "-rf" as a second filename: $ mkdir -p /tmp/rmtest/sub && rm /tmp/rmtest -rf rm: /tmp/rmtest: is a directory rm: -rf: No such file or directory exit=1 So nothing is removed, and each of the four calls prints two errors plus "system command failed (256)" from check_system. Every run of a tool that owns an OpenpilotPrefix (replay, cabana) leaks its params dir, its comma_home and its /tmp/msgq_ dir; 33 of each had accumulated on my machine. Pass the flags first. * replay: capture downloader's stderr so download progress is reported again (#38734) * bump panda (new health packet) (#38736) pandad: support compact health packet * BMRLNAP (#38681) * ui: clarify branch switcher error message (#38732) * ui(mici): name updater signal constants (#38731) * mici: name updater signal constants * drop SIGNAL_ prefix * self contained --------- Co-authored-by: Shane Smiskol <shane@smiskol.com> * modem.py: accept hex chars in ICCID (#38735) E.118 specifies decimal digits, but many real SIMs carry hex characters in EF_ICCID (e.g. China Mobile's 898600B5... range, some MVNO/IoT SIMs). AT+QCCID returns them verbatim, and the strict isdigit() check blanked the ICCID, leaving the modem daemon stuck in INITIALIZING forever and cellular dead. ModemManager parses ICCID as hex for the same reason. Verified on a comma four with a China Mobile SIM (EG916Q-GL): previously stuck retrying 'identity read incomplete', now dials and passes traffic. * TGC (#38739) * 23e6a04e-e6e5-462b-a0bb-e4088275ee43/12864 tgc * here * monitor chestnut USB in hardwared (#38741) hardwared: monitor chestnut USB independently * modeld: wait for stable chestnut (#38742) modeld: wait for stable chestnut * Revert "monitor chestnut USB in hardwared (#38741)" (#38744) This reverts commit `7d5596d5c3`. * amd warp (#38684) * modeld: fuse warp and policy TinyJit * bump tg * fix? * this simple trick... * debug 1 * bump tg * pack all * wips * fix * BIG_INTO_SMALL remove * slower * ui: show usb connection (#38745) * ui: show USB status * ui: resize USB icon * ui: classify USB device once * ui: debounce USB disconnect * cereal: log big model in drivingModelData (#38747) * ui: show one GPU status (#38748) ui: show one GPU status icon * AGNOS 19.7 (#38750) --------- Co-authored-by: Trey Moen <50057480+greatgitsby@users.noreply.github.com> Co-authored-by: Daniel Koepping <elkoled@gmail.com> Co-authored-by: Robbe Derks <robbe.derks@gmail.com> Co-authored-by: Harald Schäfer <harald.the.engineer@gmail.com> Co-authored-by: Shane Smiskol <shane@smiskol.com> Co-authored-by: XiaoXX <xiaoxx97@outlook.com> Co-authored-by: YassineYousfi <yyousfi1@binghamton.edu>	2026-09-02 13:57:07 -04:00
Jason Wen	f5bb855477	Merge commit '6249f4d5b0e63c05f08bce12ca3afebda9f764a3' into sync-20260902 # Conflicts: # openpilot/selfdrive/modeld/SConscript # openpilot/selfdrive/modeld/modeld.py # openpilot/selfdrive/pandad/pandad.cc # openpilot/selfdrive/selfdrived/alerts_offroad.json # openpilot/selfdrive/ui/layouts/onboarding.py # openpilot/selfdrive/ui/mici/layouts/home.py # openpilot/system/hardware/hardwared.py # panda # tinygrad_repo	2026-09-02 13:47:27 -04:00