This model that model sunny (#1970)

* ci: compile default models with stock modeld (#1954) * ci: compile default big model with stock modeld * Revert "Revert big RL model (#38627)" This reverts commit 516ec1e682. * ci: compile default small model with stock modeld compiler * Reapply "Revert big RL model (#38627)" This reverts commit d06cfabb62. * ci: offload small model compilation (#1952) * ci: compile default big model with stock modeld * Revert "Revert big RL model (#38627)" This reverts commit 516ec1e682. * ci: compile default small model with stock modeld compiler * ci: offload small model compilation * Reapply "Revert big RL model (#38627)" This reverts commit d06cfabb62. * Reapply "Revert big RL model (#38627)" This reverts commit d06cfabb62. * ci: fix DM model build, separate HF defaults paths, nuke build races (#1956) * ci: fix DM model build, separate HF defaults paths, nuke build races * more split! * name * ci: download driving and DM model chunks into chestnut prebuilt output * ci: parallelize models chunk downloads and split branch publishing (#1955) * ci: parallelize model chunk downloads and better publish * ci: download all model chunks in parallel with xargs -P8 * split split * ew * must require * ci: identical environment for publish_chestnut prebuilt * ci: add tinygrad ref check to prepare_chestnut and even faster prebuilt stages (#1957) * ci: faster prebuilt stages * tg check chestnut * zoomer! * modeld_v2: spatial features (#1934) * modeld_v2: spatial features * Update fetcher.py * dont reshape non 4 dim arrays * realize for non compiled * Update compile_modeld.py * god dammit it was realize() * it was fucking frozen tinygrad. just need to recompile * bump * ci: add is_big flag to metadata.json to support backward compat * Update model_generator.py * Update sunnypilot-build-model.yaml * Update helpers.py * Revert "Update helpers.py" This reverts commit 3a955ca11a. * Reapply "Update helpers.py" This reverts commit ca9c6e1933. * models: use less strict chestnut detection state --------- Co-authored-by: Jason Wen <haibin.wen3@gmail.com> * ui/models: handle missing files during cache size calculation (#1958) * models: fix current model not updating on chestnut status (#1959) * models: preserve user model selection across reboots and power cycles * no * again * idk * over * models: persist model selection per catalog across chestnut state changes (#1960) * [MICI] ui: four-state eGPU icon for non-default big models (#1945) * ui: four-state eGPU icon for non-default big models * oops * try this out * align * ui: fix scrolling label speed at non-60fps refresh rates (#1967) * ui: fix scrolling label speed at non-60fps refresh rates * send it * nope * more * [TIZI/TICI] sidebar: show eGPU icon when chestnut is present (#1968) * [tizi/tici] sidebar: show eGPU icon when chestnut is present * matchy match * fix * ui: use full big model failure detection for sidebar and home eGPU icons (#1969) * models: dual-slot backend (qcom/usbgpu) with ref-based downloads (#1966) * models: dual-slot backend (qcom/usbgpu) with ref-based downloads * models: restore get_active_source and the usbgpu-to-qcom fallback * models: fix per-slot validation and cap mismatched-source refetches * ui/models: select models by ref and seed the usbgpu slot on migration * models: drop defensive attribute guards on capnp bundles * models: remove vestigial fetcher state and dead fallbacks * models: resolve the active bundle from the active source slot only * models: pass the usbgpu kwarg through the modeld test stubs * models: resolve the displayed model from the active slot in ui_state * models: correct the validation memo type hint * models: drop docstrings that restate the function name --------- Co-authored-by: Jason Wen <haibin.wen3@gmail.com> * ui: unify model source predicate and per-source bundle lookup in model_info * [TIZI/TICI] ui: disable the other-model row onroad like the active row * [TIZI/TICI] ui: drop docstring that restates the function name * [TIZI/TICI] ui: keep Favorites as the first model folder in the picker * ui: record why model names read the params slots and not modelManagerSP * ui: show the default model's name on the picker Default entries * models: bind a download to its ref so cancel and reselect work everywhere * models: resume partial chunked downloads and verify silently * models: publish a verifying status so cached checks read as verification, not a stuck download * [TIZI/TICI] ui: move download status onto each model's own row * [TIZI/TICI] ui: show the row status description while it has text * [TIZI/TICI] ui: restore the Model Status bar row * models: a cancel interrupts verification immediately and keeps on-disk chunks * models: a selection made mid-download queues instead of cancelling the transfer * [TIZI/TICI] ui: Model Status shows both slots idle and the queued pick while busy * [TIZI/TICI] ui: label the Model Status slots small and big and scroll long names * models: start a queued download in the same tick and label empty slots (Default) * ui: scroll Model Status names at the corrected speed * [TIZI/TICI] ui: Model Status shows the big model failing over to small * [TIZI/TICI] ui: stable model rows and a runner-matched failover note on Model Status * [TIZI/TICI] ui: model rows show full names and the failover note reopens with the page * ui: name the actually driving model runner-matched and bring mici to state parity * fix ugly --------- Co-authored-by: James Vecellio-Grant <159560811+Discountchubbs@users.noreply.github.com> Co-authored-by: Nayan <nayan8teen@gmail.com>
2026-09-09 20:13:43 +08:00 · 2026-08-27 00:07:40 -04:00
parent bbb3c01b06 ae4d48bd16
commit c02beedcc6
33 changed files with 1421 additions and 386 deletions
@@ -10,13 +10,18 @@ on:
        options:
          - small
          - big
+          - dm
  workflow_call:
    inputs:
      target:
-        description: 'Model target to build (small or big)'
+        description: 'Model target to build (small, big, or dm)'
        required: true
        type: string

+concurrency:
+  group: build-default-models-${{ inputs.target }}
+  cancel-in-progress: false
+
 env:
  HF_REPO: sunnypilot/sunnypilot_models_v1

@@ -28,9 +33,7 @@ jobs:
      onnx_ref: ${{ steps.resolve.outputs.onnx_ref }}
      onnx_path: ${{ steps.resolve.outputs.onnx_path }}
      hf_defaults_path: ${{ steps.resolve.outputs.hf_defaults_path }}
-      target_hardware: ${{ steps.resolve.outputs.target_hardware }}
      tinygrad_ref: ${{ steps.resolve.outputs.tinygrad_ref }}
-      dm_onnx_ref: ${{ steps.resolve.outputs.dm_onnx_ref }}
    steps:
      - uses: actions/checkout@v4
        with:
@@ -44,12 +47,14 @@ jobs:
            NAME=$(python3 -c "from openpilot.sunnypilot.models.model_name import DEFAULT_BIG_MODEL; print(DEFAULT_BIG_MODEL)")
            ONNX_PATH="openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx"
            HF_DEFAULTS_PATH="models/defaults/big"
-            TARGET_HW="usbgpu"
+          elif [ "${{ inputs.target }}" = "dm" ]; then
+            ONNX_PATH="openpilot/selfdrive/modeld/models/dmonitoring_model.onnx"
+            HF_DEFAULTS_PATH="models/defaults/dm"
+            NAME="dmonitoring_model ($(git log -1 --format=%cd --date=format:'%B %d, %Y' -- "$ONNX_PATH"))"
          else
            NAME=$(python3 -c "from openpilot.sunnypilot.models.model_name import DEFAULT_MODEL; print(DEFAULT_MODEL)")
            ONNX_PATH="openpilot/selfdrive/modeld/models/driving_supercombo.onnx"
            HF_DEFAULTS_PATH="models/defaults/small"
-            TARGET_HW="qcom"
          fi

          ONNX_REF=$(git log -1 --format='%H' -- "$ONNX_PATH")
@@ -59,65 +64,279 @@ jobs:
            exit 1
          fi

-          DM_ONNX_REF=""
-          if [ "${{ inputs.target }}" = "small" ]; then
-            DM_ONNX_REF=$(git log -1 --format='%H' -- openpilot/selfdrive/modeld/models/dmonitoring_model.onnx)
-          fi
-
          echo "model_name=${NAME}" >> $GITHUB_OUTPUT
          echo "onnx_ref=${ONNX_REF}" >> $GITHUB_OUTPUT
          echo "onnx_path=${ONNX_PATH}" >> $GITHUB_OUTPUT
          echo "hf_defaults_path=${HF_DEFAULTS_PATH}" >> $GITHUB_OUTPUT
-          echo "target_hardware=${TARGET_HW}" >> $GITHUB_OUTPUT
          echo "tinygrad_ref=${TINYGRAD_REF}" >> $GITHUB_OUTPUT
-          echo "dm_onnx_ref=${DM_ONNX_REF}" >> $GITHUB_OUTPUT

-  build_driving_model:
+  build_small_model:
    needs: resolve
-    uses: ./.github/workflows/sunnypilot-build-model.yaml
-    with:
-      upstream_branch: ${{ needs.resolve.outputs.onnx_ref }}
-      custom_name: ${{ needs.resolve.outputs.model_name }}
-      target_hardware: ${{ needs.resolve.outputs.target_hardware }}
-    secrets: inherit
+    if: ${{ inputs.target == 'small' }}
+    runs-on: [self-hosted, tici]
+    env:
+      SMALL_ONNX: openpilot/selfdrive/modeld/models/driving_supercombo.onnx
+      SMALL_PKL: openpilot/selfdrive/modeld/models/driving_tinygrad.pkl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Pull ONNX via LFS
+        run: git lfs pull -I "${{ env.SMALL_ONNX }}"
+
+      - name: Set environment variables
+        run: |
+          source /etc/profile
+          export UV_PROJECT_ENVIRONMENT=${HOME}/venv
+          export UV_PYTHON_PREFERENCE=managed
+          export UV_PYTHON_INSTALL_DIR=${HOME}/uv/python
+          export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
+          uv sync --frozen
+          printenv >> $GITHUB_ENV
+
+      - name: Disable powersave
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          PYTHONPATH=$PYTHONPATH:${{ github.workspace }}/ ${{ github.workspace }}/scripts/manage-powersave.py --disable
+
+      - name: Compile small model with stock compiler
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH="${PYTHONPATH}:${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}"
+
+          MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
+          CAMERA_RES=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
+          FRAME_SKIP=$(python3 -c "from openpilot.selfdrive.modeld.constants import ModelConstants as MC; print(MC.MODEL_RUN_FREQ // MC.MODEL_CONTEXT_FREQ)")
+
+          TG_FLAGS="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+
+          env ${TG_FLAGS} python3 \
+            ${{ github.workspace }}/openpilot/selfdrive/modeld/compile_modeld.py \
+            --onnx ${{ github.workspace }}/${{ env.SMALL_ONNX }} \
+            --model-size $MODEL_SIZE \
+            --camera-resolutions $CAMERA_RES \
+            --frame-skip $FRAME_SKIP \
+            --output ${{ github.workspace }}/${{ env.SMALL_PKL }}
+
+      - name: Chunk small pkl
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH=${{ github.workspace }}
+          python3 -c "
+          from openpilot.common.file_chunker import chunk_file, get_chunk_targets
+          import os
+          pkl = '${{ github.workspace }}/${{ env.SMALL_PKL }}'
+          size = os.path.getsize(pkl)
+          targets = get_chunk_targets(pkl, size)
+          chunk_file(pkl, targets)
+          print(f'Chunked into {len(targets)} files')
+          "
+
+      - name: Prepare output
+        env:
+          MODEL_NAME: ${{ needs.resolve.outputs.model_name }}
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH=${{ github.workspace }}
+          MODELS_DIR="${{ github.workspace }}/openpilot/selfdrive/modeld/models"
+          OUTPUT_DIR="${{ github.workspace }}/small_output"
+          PKL_BASE="driving_tinygrad.pkl"
+          mkdir -p "$OUTPUT_DIR"
+
+          cp "$MODELS_DIR/${PKL_BASE}".chunk* "$OUTPUT_DIR/"
+          cp "$MODELS_DIR/${PKL_BASE}.chunkmanifest" "$OUTPUT_DIR/"
+
+          python3 "${{ github.workspace }}/release/ci/model_generator.py" \
+            --model-dir "$MODELS_DIR" \
+            --output-dir "$OUTPUT_DIR" \
+            --custom-name "$MODEL_NAME" \
+            --upstream-branch "${{ needs.resolve.outputs.onnx_ref }}"
+
+          echo "model-${MODEL_NAME}-${{ github.run_number }}" > "$OUTPUT_DIR/artifact_name.txt"
+
+      - name: Upload small model artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: model-${{ needs.resolve.outputs.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 }}
+          path: ${{ github.workspace }}/small_output/artifact_name.txt
+
+      - name: Re-enable powersave
+        if: always()
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          PYTHONPATH=$PYTHONPATH:${{ github.workspace }}/ ${{ github.workspace }}/scripts/manage-powersave.py --enable
+
+  build_big_model:
+    needs: resolve
+    if: ${{ inputs.target == 'big' }}
+    runs-on: [self-hosted, usbgpu]
+    env:
+      BIG_ONNX: openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx
+      BIG_PKL: openpilot/selfdrive/modeld/models/big_driving_tinygrad.pkl
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+
+      - name: Pull big ONNX via LFS
+        run: git lfs pull -I "${{ env.BIG_ONNX }}"
+
+      - name: Set environment variables
+        run: |
+          source /etc/profile
+          export UV_PROJECT_ENVIRONMENT=${HOME}/venv
+          export UV_PYTHON_PREFERENCE=managed
+          export UV_PYTHON_INSTALL_DIR=${HOME}/uv/python
+          export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
+          uv sync --frozen
+          printenv >> $GITHUB_ENV
+
+      - name: Disable powersave
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          PYTHONPATH=$PYTHONPATH:${{ github.workspace }}/ ${{ github.workspace }}/scripts/manage-powersave.py --disable
+
+      - name: Wait for chestnut PCIe link
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH="${PYTHONPATH}:${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}"
+          python3 -c "
+          import time
+          from openpilot.system.hardware.chestnut.flash import link_up
+          for i in range(10):
+              if link_up():
+                  print(f'PCIe link up after {i+1} attempt(s)')
+                  break
+              time.sleep(1)
+          else:
+              raise RuntimeError('Chestnut PCIe link not ready after 10 attempts')
+          "
+
+      - name: Compile big model with stock compiler
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH="${PYTHONPATH}:${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}"
+
+          MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
+          CAMERA_RES=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
+          FRAME_SKIP=$(python3 -c "from openpilot.selfdrive.modeld.constants import ModelConstants as MC; print(MC.MODEL_RUN_FREQ // MC.MODEL_CONTEXT_FREQ)")
+
+          TG_FLAGS="DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2"
+
+          env ${TG_FLAGS} python3 \
+            ${{ github.workspace }}/openpilot/selfdrive/modeld/compile_modeld.py \
+            --onnx ${{ github.workspace }}/${{ env.BIG_ONNX }} \
+            --model-size $MODEL_SIZE \
+            --camera-resolutions $CAMERA_RES \
+            --frame-skip $FRAME_SKIP \
+            --output ${{ github.workspace }}/${{ env.BIG_PKL }}
+
+      - name: Chunk big pkl
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH=${{ github.workspace }}
+          python3 -c "
+          from openpilot.common.file_chunker import chunk_file, get_chunk_targets
+          import os
+          pkl = '${{ github.workspace }}/${{ env.BIG_PKL }}'
+          size = os.path.getsize(pkl)
+          targets = get_chunk_targets(pkl, size)
+          chunk_file(pkl, targets)
+          print(f'Chunked into {len(targets)} files')
+          "
+
+      - name: Prepare output
+        env:
+          MODEL_NAME: ${{ needs.resolve.outputs.model_name }}
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          export PYTHONPATH=${{ github.workspace }}
+          MODELS_DIR="${{ github.workspace }}/openpilot/selfdrive/modeld/models"
+          OUTPUT_DIR="${{ github.workspace }}/big_output"
+          PKL_BASE="big_driving_tinygrad.pkl"
+          mkdir -p "$OUTPUT_DIR"
+
+          cp "$MODELS_DIR/${PKL_BASE}".chunk* "$OUTPUT_DIR/"
+          cp "$MODELS_DIR/${PKL_BASE}.chunkmanifest" "$OUTPUT_DIR/"
+
+          python3 "${{ github.workspace }}/release/ci/model_generator.py" \
+            --model-dir "$MODELS_DIR" \
+            --output-dir "$OUTPUT_DIR" \
+            --custom-name "$MODEL_NAME" \
+            --upstream-branch "${{ needs.resolve.outputs.onnx_ref }}"
+
+          echo "model-${MODEL_NAME}-${{ github.run_number }}" > "$OUTPUT_DIR/artifact_name.txt"
+
+      - name: Upload big model artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: model-${{ needs.resolve.outputs.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 }}
+          path: ${{ github.workspace }}/big_output/artifact_name.txt
+
+      - name: Re-enable powersave
+        if: always()
+        run: |
+          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          PYTHONPATH=$PYTHONPATH:${{ github.workspace }}/ ${{ github.workspace }}/scripts/manage-powersave.py --enable

  upload_defaults:
-    needs: [ resolve, build_driving_model, build_dm_model ]
-    if: ${{ !cancelled() && needs.build_driving_model.result == 'success' && (inputs.target != 'small' || needs.build_dm_model.result == 'success') }}
+    needs: [ resolve, build_small_model, build_big_model, build_dm_model ]
+    if: |
+      ${{
+        !cancelled() &&
+        (inputs.target == 'big' && needs.build_big_model.result == 'success' ||
+         inputs.target == 'small' && needs.build_small_model.result == 'success' ||
+         inputs.target == 'dm' && needs.build_dm_model.result == 'success')
+      }}
    runs-on: ubuntu-24.04
    permissions:
      id-token: write
      contents: write
-    env:
-      DM_ONNX: openpilot/selfdrive/modeld/models/dmonitoring_model.onnx
    steps:
      - uses: actions/checkout@v4

      - name: Pull ONNX via LFS
-        run: git lfs pull -I "${{ needs.resolve.outputs.onnx_path }}${{ inputs.target == 'small' && ',openpilot/selfdrive/modeld/models/dmonitoring_model.onnx' || '' }}"
+        run: git lfs pull -I "${{ needs.resolve.outputs.onnx_path }}"

      - name: Install huggingface_hub
        run: pip install --upgrade "huggingface_hub>=0.22.0"

-      - name: Download driving artifact name
+      - name: Download artifact name
+        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        uses: actions/download-artifact@v4
        with:
          name: artifact-name-${{ needs.resolve.outputs.model_name }}
          path: artifact_name

-      - name: Read driving artifact name
+      - name: Read artifact name
+        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        id: artifact
        run: |
          ARTIFACT_NAME=$(cat artifact_name/artifact_name.txt)
          echo "artifact_name=$ARTIFACT_NAME" >> $GITHUB_OUTPUT

-      - name: Download driving model artifact
+      - name: Download model artifact
+        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        uses: actions/download-artifact@v4
        with:
          name: ${{ steps.artifact.outputs.artifact_name }}
          path: output

-      - name: Upload driving model to HF
+      - name: Upload model to HF
+        if: ${{ inputs.target == 'small' || inputs.target == 'big' }}
        env:
          HF_OIDC_RESOURCE: datasets/${{ env.HF_REPO }}
          ARTIFACT_NAME: ${{ steps.artifact.outputs.artifact_name }}
@@ -136,14 +355,14 @@ jobs:
            --run-number "${{ github.run_number }}"

      - name: Download DM artifact
-        if: ${{ inputs.target == 'small' }}
+        if: ${{ inputs.target == 'dm' }}
        uses: actions/download-artifact@v4
        with:
          name: dm-model-${{ github.run_number }}
          path: dm_output

      - name: Generate DM metadata and upload to HF
-        if: ${{ inputs.target == 'small' }}
+        if: ${{ inputs.target == 'dm' }}
        env:
          HF_OIDC_RESOURCE: datasets/${{ env.HF_REPO }}
        run: |
@@ -176,8 +395,8 @@ jobs:
          metadata = {
              'bundles': [{
                  'short_name': 'DMMODEL',
-                  'display_name': 'dmonitoring_model',
-                  'ref': '${{ needs.resolve.outputs.dm_onnx_ref }}',
+                  'display_name': '${{ needs.resolve.outputs.model_name }}',
+                  'ref': '${{ needs.resolve.outputs.onnx_ref }}',
                  'runner': 'tinygrad',
                  'build_time': datetime.now(UTC).strftime('%Y-%m-%dT%H:%M:%SZ'),
                  'models': [{
@@ -200,15 +419,15 @@ jobs:
            --hf-defaults-path "${{ needs.resolve.outputs.hf_defaults_path }}" \
            --artifact-name "dm-model-${{ github.run_number }}" \
            --model-dir dm_output \
-            --onnx-path "${{ env.DM_ONNX }}" \
-            --onnx-ref "${{ needs.resolve.outputs.dm_onnx_ref }}" \
-            --model-name "dmonitoring_model" \
+            --onnx-path "${{ needs.resolve.outputs.onnx_path }}" \
+            --onnx-ref "${{ needs.resolve.outputs.onnx_ref }}" \
+            --model-name "${{ needs.resolve.outputs.model_name }}" \
            --tinygrad-ref "${{ needs.resolve.outputs.tinygrad_ref }}" \
            --run-number "${{ github.run_number }}"

  build_dm_model:
    needs: resolve
-    if: ${{ inputs.target == 'small' }}
+    if: ${{ inputs.target == 'dm' }}
    runs-on: [self-hosted, tici]
    env:
      DM_ONNX: openpilot/selfdrive/modeld/models/dmonitoring_model.onnx
@@ -218,6 +437,9 @@ jobs:
        with:
          submodules: recursive

+      - name: Pull DM ONNX via LFS
+        run: git lfs pull -I "${{ env.DM_ONNX }}"
+
      - name: Set environment variables
        run: |
          source /etc/profile
@@ -0,0 +1,66 @@
+name: Download HF model chunks
+description: Resolve and download model chunks from HuggingFace in parallel
+
+inputs:
+  hf_repo:
+    description: HuggingFace dataset repo
+    required: true
+  models:
+    description: 'JSON array of {hf_path, onnx_hash, canonical} objects'
+    required: true
+  dest_dir:
+    description: Destination directory for downloaded chunks
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Download model chunks
+      shell: bash
+      env:
+        HF_REPO: ${{ inputs.hf_repo }}
+        MODELS_JSON: ${{ inputs.models }}
+        DEST_DIR: ${{ inputs.dest_dir }}
+      run: |
+        set -eo pipefail
+        DOWNLOAD_LIST=$(mktemp)
+
+        resolve_chunks() {
+          local HF_PATH="$1" ONNX_HASH="$2" CANONICAL="$3" DEST_DIR="$4"
+          local JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_PATH}/default_models.json"
+          local DEFAULTS BUNDLE ARTIFACT BASE_URL NUM_CHUNKS
+          DEFAULTS=$(curl -fsSL "$JSON_URL")
+          BUNDLE=$(echo "$DEFAULTS" | jq --arg hash "$ONNX_HASH" '.bundles[] | select(.onnx_sha256 == $hash)')
+          ARTIFACT=$(echo "$BUNDLE" | jq -r '.models[0].artifact')
+          BASE_URL=$(echo "$ARTIFACT" | jq -r '.download_uri.url' | sed 's|/[^/]*$||')
+          NUM_CHUNKS=$(echo "$ARTIFACT" | jq -r '.chunks | length')
+
+          mkdir -p "$DEST_DIR"
+          while IFS= read -r CHUNK_NAME; do
+            CHUNK_IDX=$(echo "$CHUNK_NAME" | grep -oP 'chunk\K[0-9]+of[0-9]+' || true)
+            if [ -z "$CHUNK_IDX" ]; then
+              echo "::error::Failed to parse chunk index from: $CHUNK_NAME"
+              return 1
+            fi
+            ENCODED_URL=$(python3 -c "import urllib.parse; print(urllib.parse.quote('${BASE_URL}/${CHUNK_NAME}', safe=':/'))")
+            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}/${CANONICAL}.chunkmanifest"
+        }
+
+        echo "$MODELS_JSON" | jq -c '.[]' | while IFS= read -r model; do
+          HF_PATH=$(echo "$model" | jq -r '.hf_path')
+          ONNX_HASH=$(echo "$model" | jq -r '.onnx_hash')
+          CANONICAL=$(echo "$model" | jq -r '.canonical')
+          resolve_chunks "$HF_PATH" "$ONNX_HASH" "$CANONICAL" "$DEST_DIR"
+        done
+
+        TOTAL=$(wc -l < "$DOWNLOAD_LIST")
+        echo "Downloading $TOTAL chunks with 8 parallel connections..."
+        xargs -P8 -d'\n' -I{} bash -c '
+          URL="${1%%	*}"
+          DEST="${1#*	}"
+          echo "Downloading $(basename "$DEST")"
+          curl -fsSL --retry 3 --retry-delay 5 -o "$DEST" "$URL"
+        ' _ {} < "$DOWNLOAD_LIST"
+        rm -f "$DOWNLOAD_LIST"
@@ -188,7 +188,7 @@ jobs:
          if [ "${{ inputs.target_hardware }}" == "usbgpu" ]; then
            echo "USBGPU build"
            export USBGPU=1
-            TG_FLAGS="DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2"
+            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2"
            OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
          else
            echo "QCOM build"
@@ -39,6 +39,8 @@ jobs:
      include_big_model: ${{ steps.strategy.outputs.include_big_model }}
    steps:
      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
      - name: Extract deploy strategy
        id: strategy
        run: |
@@ -96,6 +98,8 @@ jobs:
      }}
    steps:
      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
      - name: Wait for Tests
        uses: ./.github/workflows/wait-for-action # Path to where you place the action
        with:
@@ -119,6 +123,7 @@ jobs:
    steps:
      - uses: actions/checkout@v4
        with:
+          fetch-depth: 1
          submodules: recursive
          ref: ${{ env.SOURCE_BRANCH }}
          repository: ${{ github.event.pull_request.head.repo.fork && github.event.pull_request.head.repo.full_name || github.repository }}
@@ -165,7 +170,7 @@ jobs:
          scons -j1 cache_dir="$SCONS_CACHE" --minimal \
            openpilot/selfdrive/locationd openpilot/sunnypilot/selfdrive/locationd
          echo "Building rest of sunnypilot"
-          /usr/bin/time -v scons -j$(nproc) cache_dir="$SCONS_CACHE" --minimal
+          SKIP_TINYGRAD_COMPILE=1 /usr/bin/time -v scons -j$(nproc) cache_dir="$SCONS_CACHE" --minimal
          touch ${BUILD_DIR}/prebuilt
          if [[ "${{ runner.debug }}" == "1" ]]; then
            ls -la ${BUILD_DIR}
@@ -214,59 +219,174 @@ jobs:
    outputs:
      onnx_sha256: ${{ steps.resolve.outputs.onnx_sha256 }}
    env:
+      GH_REPO: ${{ github.repository }}
      HF_REPO: sunnypilot/sunnypilot_models_v1
      HF_DEFAULTS_PATH: models/defaults/big
    steps:
-      - uses: actions/checkout@v4
-        with:
-          ref: ${{ github.head_ref || github.ref_name }}
-
-      - run: git lfs pull -I "openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx"
-
-      - name: Check HF defaults and build if needed
+      - name: Resolve ONNX hash and tinygrad ref via API
        id: resolve
        run: |
-          ACTUAL_ONNX_HASH=$(sha256sum "openpilot/selfdrive/modeld/models/big_driving_supercombo.onnx" | cut -d' ' -f1)
-          echo "Repo ONNX hash: $ACTUAL_ONNX_HASH"
-          echo "onnx_sha256=$ACTUAL_ONNX_HASH" >> $GITHUB_OUTPUT
+          REF="${{ 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)
+          echo "ONNX hash: $ONNX_HASH"
+          echo "onnx_sha256=$ONNX_HASH" >> $GITHUB_OUTPUT
+
+          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
+          echo "tinygrad ref: $TINYGRAD_REF"

          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"

-          check_hash() {
+          check_defaults() {
            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
-            BUNDLE=$(echo "$DEFAULTS" | jq --arg hash "$ACTUAL_ONNX_HASH" '.bundles[] | select(.onnx_sha256 == $hash)' 2>/dev/null)
+            TINYGRAD_MATCH=$(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 == $hash)' 2>/dev/null)
            [ -n "$BUNDLE" ] && [ "$BUNDLE" != "null" ]
          }

-          if check_hash; then
-            echo "HF defaults match repo ONNX"
-          else
-            echo "No matching model on HF — triggering build"
-            gh workflow run build-default-models.yaml --ref "${{ github.head_ref || github.ref_name }}" -f target=big
-
-            echo "Waiting for build to start..."
-            sleep 120
-
-            RUN_ID=$(gh run list --workflow=build-default-models.yaml --branch="${{ github.head_ref || github.ref_name }}" --limit=1 --json databaseId --jq '.[0].databaseId')
-            if [ -z "$RUN_ID" ] || [ "$RUN_ID" = "null" ]; then
-              echo "::error::Failed to find build-default-models run"
-              exit 1
-            fi
-
-            echo "Waiting for run $RUN_ID..."
-            gh run watch "$RUN_ID"
-
-            CONCLUSION=$(gh run view "$RUN_ID" --json conclusion --jq '.conclusion')
-            if [ "$CONCLUSION" != "success" ]; then
-              echo "::error::build-default-models failed: $CONCLUSION"
-              exit 1
-            fi
-
-            if ! check_hash; then
-              echo "::error::HF defaults still don't match after build"
-              exit 1
-            fi
+          if check_defaults; then
+            echo "HF defaults match repo ONNX hash and tinygrad ref"
+            exit 0
          fi
+
+          echo "No matching model on HF — dispatching build"
+          gh workflow run build-default-models.yaml --ref "$REF" -f target=big
+
+          echo "Polling HF for big model availability..."
+          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
+            fi
+            echo "Poll $i/90: not yet available"
+          done
+
+          echo "::error::Big model not available on HF after 45 minutes"
+          exit 1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Cancel run on failure
+        if: failure()
+        run: gh run cancel ${{ github.run_id }}
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  prepare_small_model:
+    needs: [ prepare_strategy ]
+    runs-on: ubuntu-24.04
+    outputs:
+      driving_onnx_sha256: ${{ steps.resolve.outputs.driving_onnx_sha256 }}
+    env:
+      GH_REPO: ${{ github.repository }}
+      HF_REPO: sunnypilot/sunnypilot_models_v1
+      HF_DEFAULTS_PATH: models/defaults/small
+    steps:
+      - name: Resolve ONNX hash and tinygrad ref via API
+        id: resolve
+        run: |
+          REF="${{ 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)
+          echo "Driving ONNX hash: $DRIVING_HASH"
+          echo "driving_onnx_sha256=$DRIVING_HASH" >> $GITHUB_OUTPUT
+
+          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
+          echo "tinygrad ref: $TINYGRAD_REF"
+
+          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"
+
+          check_defaults() {
+            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
+            TINYGRAD_MATCH=$(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 == $hash)' 2>/dev/null)
+            [ -n "$DRIVING" ] && [ "$DRIVING" != "null" ] || return 1
+          }
+
+          if check_defaults; then
+            echo "HF defaults match repo ONNX hash and tinygrad ref"
+            exit 0
+          fi
+
+          echo "No matching model on HF — dispatching build"
+          gh workflow run build-default-models.yaml --ref "$REF" -f target=small
+
+          echo "Polling HF for model availability..."
+          for i in $(seq 1 60); do
+            sleep 30
+            if check_defaults; then
+              echo "Model available on HF after $((i * 30))s"
+              exit 0
+            fi
+            echo "Poll $i/60: not yet available"
+          done
+
+          echo "::error::Small driving model not available on HF after 30 minutes"
+          exit 1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Cancel run on failure
+        if: failure()
+        run: gh run cancel ${{ github.run_id }}
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+
+  prepare_dm_model:
+    needs: [ prepare_strategy ]
+    runs-on: ubuntu-24.04
+    outputs:
+      dm_onnx_sha256: ${{ steps.resolve.outputs.dm_onnx_sha256 }}
+    env:
+      GH_REPO: ${{ github.repository }}
+      HF_REPO: sunnypilot/sunnypilot_models_v1
+      HF_DEFAULTS_PATH: models/defaults/dm
+    steps:
+      - name: Resolve ONNX hash and tinygrad ref via API
+        id: resolve
+        run: |
+          REF="${{ 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)
+          echo "DM ONNX hash: $DM_HASH"
+          echo "dm_onnx_sha256=$DM_HASH" >> $GITHUB_OUTPUT
+
+          TINYGRAD_REF=$(gh api "repos/${GH_REPO}/contents/tinygrad_repo?ref=${REF}" --jq '.sha')
+          echo "tinygrad ref: $TINYGRAD_REF"
+
+          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"
+
+          check_defaults() {
+            DEFAULTS=$(curl -fsSL "$JSON_URL" 2>/dev/null) || return 1
+            TINYGRAD_MATCH=$(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 == $hash)' 2>/dev/null)
+            [ -n "$DM" ] && [ "$DM" != "null" ] || return 1
+          }
+
+          if check_defaults; then
+            echo "HF defaults match DM ONNX hash and tinygrad ref"
+            exit 0
+          fi
+
+          echo "No matching DM model on HF — dispatching build"
+          gh workflow run build-default-models.yaml --ref "$REF" -f target=dm
+
+          echo "Polling HF for DM model availability..."
+          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
+            fi
+            echo "Poll $i/60: not yet available"
+          done
+
+          echo "::error::DM model not available on HF after 30 minutes"
+          exit 1
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

@@ -278,23 +398,24 @@ jobs:

  publish:
    concurrency:
-      # We do a bit of a hack here to avoid canceling the publishing job if a new commit comes in while we're publishing by adding the sha to the group name.
-      # This means that if multiple commits come in while we're publishing, they will be queued up and publish one after the other.
-      # Otherwise, if a job is waiting to be published due to environment wait time, it would be canceled by a new commit and restart the wait time.
      group: ${{ needs.prepare_strategy.outputs.publish_concurrency_group }}
      cancel-in-progress: ${{ needs.prepare_strategy.outputs.cancel_publish_in_progress == 'true' }}
    if: ${{
      always() && !cancelled() &&
      needs.build.result == 'success' &&
      needs.prepare_strategy.result == 'success' &&
+      needs.prepare_small_model.result == 'success' &&
+      needs.prepare_dm_model.result == 'success' &&
      (!contains(github.event_name, 'pull_request') || (github.event.action == 'labeled' && github.event.label.name == 'prebuilt')) &&
      (needs.prepare_strategy.outputs.include_big_model != 'true' || needs.prepare_chestnut.result == 'success')
      }}
-    needs: [ build, prepare_strategy, prepare_chestnut ]
+    needs: [ build, prepare_strategy, prepare_chestnut, prepare_small_model, prepare_dm_model ]
    runs-on: ubuntu-24.04
    environment: ${{ needs.prepare_strategy.outputs.environment }}
    steps:
      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1

      - name: Download prebuilt artifact
        uses: actions/download-artifact@v4
@@ -306,43 +427,16 @@ jobs:
          mkdir -p ${{ env.OUTPUT_DIR }}
          tar xzf prebuilt.tar.gz -C ${{ env.OUTPUT_DIR }}

-      - name: Prepare chestnut output
-        if: ${{ needs.prepare_chestnut.result == 'success' }}
-        run: |
-          mkdir -p "${{ github.workspace }}/chestnut_output"
-          tar xzf prebuilt.tar.gz -C "${{ github.workspace }}/chestnut_output"
-
-      - name: Download big model chunks from HF
-        if: ${{ needs.prepare_chestnut.result == 'success' }}
-        env:
-          HF_REPO: sunnypilot/sunnypilot_models_v1
-          HF_DEFAULTS_PATH: models/defaults/big
-        run: |
-          ONNX_HASH="${{ needs.prepare_chestnut.outputs.onnx_sha256 }}"
-          JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_DEFAULTS_PATH}/default_models.json"
-          DEFAULTS=$(curl -fsSL "$JSON_URL")
-          BUNDLE=$(echo "$DEFAULTS" | jq --arg hash "$ONNX_HASH" '.bundles[] | select(.onnx_sha256 == $hash)')
-
-          mkdir -p big_model_chunks
-          ARTIFACT=$(echo "$BUNDLE" | jq -r '.models[0].artifact')
-          BASE_URL=$(echo "$ARTIFACT" | jq -r '.download_uri.url' | sed 's|/[^/]*$||')
-          NUM_CHUNKS=$(echo "$ARTIFACT" | jq -r '.chunks | length')
-
-          CANONICAL="big_driving_tinygrad.pkl"
-          echo "$ARTIFACT" | jq -r '.chunks[].file_name' | while read CHUNK_NAME; do
-            CHUNK_IDX=$(echo "$CHUNK_NAME" | grep -oP 'chunk\K[0-9]+of[0-9]+')
-            CANONICAL_CHUNK="${CANONICAL}.chunk${CHUNK_IDX}"
-            ENCODED_URL=$(python3 -c "import urllib.parse; print(urllib.parse.quote('${BASE_URL}/${CHUNK_NAME}', safe=':/'))")
-            echo "Downloading $CHUNK_NAME -> $CANONICAL_CHUNK"
-            curl -fsSL -o "big_model_chunks/${CANONICAL_CHUNK}" "$ENCODED_URL"
-          done
-
-          echo "$NUM_CHUNKS" > "big_model_chunks/${CANONICAL}.chunkmanifest"
-
-      - name: Inject big model into chestnut
-        if: ${{ needs.prepare_chestnut.result == 'success' }}
-        run: |
-          cp big_model_chunks/* "${{ github.workspace }}/chestnut_output/openpilot/selfdrive/modeld/models/"
+      - name: Download model chunks from HF
+        uses: ./.github/workflows/download-hf-model-chunks
+        with:
+          hf_repo: sunnypilot/sunnypilot_models_v1
+          dest_dir: ${{ env.OUTPUT_DIR }}/openpilot/selfdrive/modeld/models
+          models: |
+            [
+              {"hf_path": "models/defaults/small", "onnx_hash": "${{ needs.prepare_small_model.outputs.driving_onnx_sha256 }}", "canonical": "driving_tinygrad.pkl"},
+              {"hf_path": "models/defaults/dm", "onnx_hash": "${{ needs.prepare_dm_model.outputs.dm_onnx_sha256 }}", "canonical": "dmonitoring_model_tinygrad.pkl"}
+            ]

      - name: Configure Git
        run: |
@@ -364,22 +458,6 @@ jobs:
            "https://x-access-token:${{github.token}}@github.com/sunnypilot/sunnypilot.git" \
            "${{ needs.prepare_strategy.outputs.extra_version_identifier }}"

-      - name: Publish chestnut branch
-        if: ${{ needs.prepare_chestnut.result == 'success' }}
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          CHESTNUT_BRANCH="${{ needs.prepare_strategy.outputs.new_branch }}-chestnut"
-          CHESTNUT_DIR="${{ github.workspace }}/chestnut_output"
-
-          ${{ env.CI_DIR }}/publish.sh \
-            "${{ github.workspace }}" \
-            "$CHESTNUT_DIR" \
-            "$CHESTNUT_BRANCH" \
-            "${{ needs.prepare_strategy.outputs.version }}" \
-            "https://x-access-token:${{github.token}}@github.com/sunnypilot/sunnypilot.git" \
-            "${{ needs.prepare_strategy.outputs.extra_version_identifier }}"
-
      - name: Tag ${{ needs.prepare_strategy.outputs.environment }}
        if: ${{ needs.prepare_strategy.outputs.is_stable_branch == 'true' && (github.event_name != 'push' || !startsWith(github.ref, 'refs/tags/')) }}
        run: |
@@ -387,12 +465,77 @@ jobs:
            git tag -f -a ${TAG} -m "${{ needs.prepare_strategy.outputs.environment }} @ ${{ needs.prepare_strategy.outputs.version }} of build ${{ needs.prepare_strategy.outputs.build }}."
            git push -f origin ${TAG}

+  publish_chestnut:
+    concurrency:
+      group: ${{ needs.prepare_strategy.outputs.publish_concurrency_group }}-chestnut
+      cancel-in-progress: ${{ needs.prepare_strategy.outputs.cancel_publish_in_progress == 'true' }}
+    if: ${{
+      always() && !cancelled() &&
+      needs.build.result == 'success' &&
+      needs.prepare_strategy.result == 'success' &&
+      needs.prepare_small_model.result == 'success' &&
+      needs.prepare_dm_model.result == 'success' &&
+      needs.prepare_chestnut.result == 'success' &&
+      (!contains(github.event_name, 'pull_request') || (github.event.action == 'labeled' && github.event.label.name == 'prebuilt'))
+      }}
+    needs: [ build, prepare_strategy, prepare_chestnut, prepare_small_model, prepare_dm_model ]
+    runs-on: ubuntu-24.04
+    environment: ${{ needs.prepare_strategy.outputs.environment }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
+
+      - name: Download prebuilt artifact
+        uses: actions/download-artifact@v4
+        with:
+          name: prebuilt
+
+      - name: Untar prebuilt
+        run: |
+          mkdir -p ${{ env.OUTPUT_DIR }}
+          tar xzf prebuilt.tar.gz -C ${{ env.OUTPUT_DIR }}
+
+      - name: Download model chunks from HF
+        uses: ./.github/workflows/download-hf-model-chunks
+        with:
+          hf_repo: sunnypilot/sunnypilot_models_v1
+          dest_dir: ${{ env.OUTPUT_DIR }}/openpilot/selfdrive/modeld/models
+          models: |
+            [
+              {"hf_path": "models/defaults/small", "onnx_hash": "${{ needs.prepare_small_model.outputs.driving_onnx_sha256 }}", "canonical": "driving_tinygrad.pkl"},
+              {"hf_path": "models/defaults/dm", "onnx_hash": "${{ needs.prepare_dm_model.outputs.dm_onnx_sha256 }}", "canonical": "dmonitoring_model_tinygrad.pkl"},
+              {"hf_path": "models/defaults/big", "onnx_hash": "${{ needs.prepare_chestnut.outputs.onnx_sha256 }}", "canonical": "big_driving_tinygrad.pkl"}
+            ]
+
+      - name: Configure Git
+        run: |
+          git config --global user.email "github-actions[bot]@users.noreply.github.com"
+          git config --global user.name "github-actions[bot]"
+
+      - name: Publish chestnut branch
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          CHESTNUT_BRANCH="${{ needs.prepare_strategy.outputs.new_branch }}-chestnut"
+
+          ${{ env.CI_DIR }}/publish.sh \
+            "${{ github.workspace }}" \
+            "${{ env.OUTPUT_DIR }}" \
+            "$CHESTNUT_BRANCH" \
+            "${{ needs.prepare_strategy.outputs.version }}" \
+            "https://x-access-token:${{github.token}}@github.com/sunnypilot/sunnypilot.git" \
+            "${{ needs.prepare_strategy.outputs.extra_version_identifier }}"
+
  notify:
    needs:
    - prepare_strategy
    - build
    - publish
+    - publish_chestnut
    - prepare_chestnut
+    - prepare_small_model
+    - prepare_dm_model
    runs-on: ubuntu-24.04
    if: ${{ (always() && !cancelled() && !failure())
      && needs.publish.result == 'success'
@@ -400,6 +543,8 @@ jobs:
      && (fromJSON(vars.DEV_FEEDBACK_NOTIFICATION_BRANCHES_V2)[github.head_ref || github.ref_name] != null) }}
    steps:
      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 1

      - name: Prepare notification message
        id: message
@@ -131,6 +131,7 @@ struct ModelManagerSP @0xaedffd8f31e7b55d {
    downloaded @2;
    cached @3;
    failed @4;
+    verifying @5;
  }

  struct DownloadProgress {
@@ -73,44 +73,45 @@ compile_modeld_script = [
 model_w, model_h = MEDMODEL_INPUT_SIZE
 frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ

-for usbgpu in [False, True] if USBGPU else [False]:
-  target_pkl_path = File(modeld_pkl_path(usbgpu)).abspath
-  # BIG_INTO_SMALL=1 builds the default target from the big model, e.g. to test it without a USB GPU
-  file_prefix, cmd_flags = ('big_', usbgpu_tg_flags) if usbgpu else ('big_' if os.getenv('BIG_INTO_SMALL') else '', tg_flags)
-  driving_onnx_deps = get_existing_chunks(File(f"models/{file_prefix}driving_supercombo.onnx").abspath)
-  camera_res_args = ' '.join(f'{cw}x{ch}' for cw, ch in CAMERA_CONFIGS)
-  # CPU 7 is isolated with isolcpus on AGNOS, so explicitly pin the compiler to it.
-  taskset = 'taskset -c 7 ' if arch == 'comma_arm64' else ''
-  cmd = (f'{cmd_flags} {mac_brew_string} {taskset}python3 {modeld_dir}/compile_modeld.py '
-        f'--model-size {model_w}x{model_h} '
-        f'--camera-resolutions {camera_res_args} '
-        f'--onnx {File(f"models/{file_prefix}driving_supercombo.onnx").abspath} '
-        f'--output {target_pkl_path} --frame-skip {frame_skip}')
-  onnx_sizes_sum = sum(os.path.getsize(f) for f in driving_onnx_deps)
-  chunk_targets = get_chunk_targets(target_pkl_path, estimate_pickle_max_size(onnx_sizes_sum))
-  def do_compile(target, source, env, command=cmd, pkl=target_pkl_path, chunks=chunk_targets):
-    from openpilot.system.hardware.chestnut.flash import link_up
-    # chestnut can enumerate before its PCIe link is up due to varying 12V power behavior across cars
-    for _ in range(10):
-      if link_up():
-        break
-      time.sleep(1)
-    else:
-      print("Chestnut not ready, skipping big model build")
-      return
-    if ret := env.Execute(command):
-      return ret
-    chunk_file(pkl, chunks)
-  def do_chunk(target, source, env, pkl=target_pkl_path, chunks=chunk_targets):
-    chunk_file(pkl, chunks)
-  actions = Action(do_compile, "  [USBGPU] $TARGET") if usbgpu else [cmd, Action(do_chunk, "  [CHUNK] $TARGET")]
-  node = lenv.Command(
-    chunk_targets,
-    tinygrad_files + compile_modeld_script + driving_onnx_deps + [Value(chunk_targets), chunker_file],
-    actions,
-  )
-  if usbgpu:
-    lenv.SideEffect(usbgpu_lock, node)
+if not os.getenv('SKIP_TINYGRAD_COMPILE'):
+  for usbgpu in [False, True] if USBGPU else [False]:
+    target_pkl_path = File(modeld_pkl_path(usbgpu)).abspath
+    # BIG_INTO_SMALL=1 builds the default target from the big model, e.g. to test it without a USB GPU
+    file_prefix, cmd_flags = ('big_', usbgpu_tg_flags) if usbgpu else ('big_' if os.getenv('BIG_INTO_SMALL') else '', tg_flags)
+    driving_onnx_deps = get_existing_chunks(File(f"models/{file_prefix}driving_supercombo.onnx").abspath)
+    camera_res_args = ' '.join(f'{cw}x{ch}' for cw, ch in CAMERA_CONFIGS)
+    # CPU 7 is isolated with isolcpus on AGNOS, so explicitly pin the compiler to it.
+    taskset = 'taskset -c 7 ' if arch == 'comma_arm64' else ''
+    cmd = (f'{cmd_flags} {mac_brew_string} {taskset}python3 {modeld_dir}/compile_modeld.py '
+          f'--model-size {model_w}x{model_h} '
+          f'--camera-resolutions {camera_res_args} '
+          f'--onnx {File(f"models/{file_prefix}driving_supercombo.onnx").abspath} '
+          f'--output {target_pkl_path} --frame-skip {frame_skip}')
+    onnx_sizes_sum = sum(os.path.getsize(f) for f in driving_onnx_deps)
+    chunk_targets = get_chunk_targets(target_pkl_path, estimate_pickle_max_size(onnx_sizes_sum))
+    def do_compile(target, source, env, command=cmd, pkl=target_pkl_path, chunks=chunk_targets):
+      from openpilot.system.hardware.chestnut.flash import link_up
+      # chestnut can enumerate before its PCIe link is up due to varying 12V power behavior across cars
+      for _ in range(10):
+        if link_up():
+          break
+        time.sleep(1)
+      else:
+        print("Chestnut not ready, skipping big model build")
+        return
+      if ret := env.Execute(command):
+        return ret
+      chunk_file(pkl, chunks)
+    def do_chunk(target, source, env, pkl=target_pkl_path, chunks=chunk_targets):
+      chunk_file(pkl, chunks)
+    actions = Action(do_compile, "  [USBGPU] $TARGET") if usbgpu else [cmd, Action(do_chunk, "  [CHUNK] $TARGET")]
+    node = lenv.Command(
+      chunk_targets,
+      tinygrad_files + compile_modeld_script + driving_onnx_deps + [Value(chunk_targets), chunker_file],
+      actions,
+    )
+    if usbgpu:
+      lenv.SideEffect(usbgpu_lock, node)

 # get model metadata
 fn = File(f"models/dmonitoring_model").abspath
@@ -142,4 +143,5 @@ def tg_compile(flags, model_name):
     Action(do_chunk, "  [CHUNK] $TARGET")],
  )

-tg_compile(tg_flags, 'dmonitoring_model')
+if not os.getenv('SKIP_TINYGRAD_COMPILE'):
+  tg_compile(tg_flags, 'dmonitoring_model')
@@ -168,9 +168,16 @@ class Sidebar(Widget, SidebarSP):
    # Home/Flag button
    flag_pressed = mouse_down and rl.check_collision_point_rec(mouse_pos, HOME_BTN)
    button_img = self._flag_img if ui_state.started else self._home_img
+    button_pos = rl.Vector2(HOME_BTN.x, HOME_BTN.y)
+    icon_opacity = 1.0
+
+    if gui_app.sunnypilot_ui():
+      button_img, button_pos, icon_opacity = SidebarSP._get_home_icon(self, button_img)

    tint = Colors.BUTTON_PRESSED if (ui_state.started and flag_pressed) else Colors.BUTTON_NORMAL
-    rl.draw_texture_ex(button_img, rl.Vector2(HOME_BTN.x, HOME_BTN.y), 0.0, 1.0, tint)
+    if icon_opacity < 1.0:
+      tint = rl.Color(tint[0], tint[1], tint[2], int(255 * icon_opacity))
+    rl.draw_texture_ex(button_img, button_pos, 0.0, 1.0, tint)

    # Microphone button
    if self._recording_audio:
@@ -248,8 +248,11 @@ class MiciHomeLayout(Widget):

    # ***** Center-aligned bottom section icons *****
    self._experimental_icon.set_visible(ui_state.experimental_mode)
-    self._egpu_icon.set_visible(ui_state.sm["deviceState"].chestnutPresent and ui_state.usbgpu_compiled)
-    self._egpu_icon_gray.set_visible(ui_state.sm["deviceState"].chestnutPresent and not ui_state.usbgpu_compiled)
+    if gui_app.sunnypilot_ui():
+      self._set_egpu_visibility()
+    else:
+      self._egpu_icon.set_visible(ui_state.sm["deviceState"].chestnutPresent and ui_state.usbgpu_compiled)
+      self._egpu_icon_gray.set_visible(ui_state.sm["deviceState"].chestnutPresent and not ui_state.usbgpu_compiled)
    self._mic_icon.set_visible(ui_state.recording_audio)
    self._body_icon.set_visible(bool(ui_state.is_body))

@@ -10,11 +10,10 @@ import time
 import pyray as rl

 from openpilot.cereal import custom
-from openpilot.sunnypilot.models.fetcher import ModelFetcher, get_cached_bundles
 from openpilot.sunnypilot.models.helpers import ACTIVE_BUNDLE_KEYS, get_selected_bundle, resolve_bundle_by_ref
 from openpilot.common.constants import CV
 from openpilot.selfdrive.ui.ui_state import device, ui_state
-from openpilot.selfdrive.ui.sunnypilot.model_info import model_info
+from openpilot.selfdrive.ui.sunnypilot.model_info import big_model_state, bundles_for_source, carrying_model, default_model_name, queued_name
 from openpilot.system.ui.lib.multilang import tr
 from openpilot.system.ui.lib.application import gui_app
 from openpilot.system.ui.widgets import DialogResult, Widget
@@ -40,6 +39,8 @@ class ModelsLayout(Widget):
    self.model_dialog = None
    self._selection_source = None
    self._downloading = False
+    self._verifying = False
+    self._last_note = None
    self.last_cache_calc_time = 0

    self._initialize_items()
@@ -51,17 +52,17 @@ class ModelsLayout(Widget):
    self._scroller = Scroller(self.items, line_separator=True, spacing=0)

  def _initialize_items(self):
-    self.current_model_item = ListItemSP(
-      title=tr("Active Model"),
+    self.small_model_item = ListItemSP(
+      title=tr("Small Model"),
      description="",
      action_item=ScrollingButtonAction(tr("SELECT")),
-      callback=self._handle_current_model_clicked
+      callback=lambda: self._open_source_dialog("qcom")
    )

-    self.other_model_item = ListItemSP(
+    self.big_model_item = ListItemSP(
      title=tr("Big Model"),
      action_item=ScrollingButtonAction(tr("SELECT")),
-      callback=self._handle_other_model_clicked
+      callback=lambda: self._open_source_dialog("usbgpu")
    )

    self.download_item = download_status_item(lambda: tr("Download") if self._downloading else tr("Model Status"))
@@ -78,7 +79,8 @@ class ModelsLayout(Widget):
      callback=self._clear_cache
    )

-    self.cancel_download_item = button_item(tr("Cancel Download"), tr("Cancel"), "",
+    self.cancel_download_item = button_item(lambda: tr("Cancel Verification") if self._verifying else tr("Cancel Download"),
+                                            tr("Cancel"), "",
                                            lambda: ui_state.params.remove("ModelManager_DownloadRef"))

    self.lane_turn_value_control = option_item_sp(tr("Adjust Lane Turn Speed"), "LaneTurnValue", 500, 2000,
@@ -104,7 +106,7 @@ class ModelsLayout(Widget):
                                        1, None, True, "", style.BUTTON_ACTION_WIDTH, None, True,
                                        lambda v: f"{v / 100:.2f} m")

-    self.items = [self.current_model_item, self.other_model_item, self.cancel_download_item, self.download_item, self.refresh_item, self.clear_cache_item,
+    self.items = [self.small_model_item, self.big_model_item, self.cancel_download_item, self.download_item, self.refresh_item, self.clear_cache_item,
                  self.lane_turn_desire_toggle, self.lane_turn_value_control, self.lagd_toggle, self.delay_control, self.camera_offset]

  def _update_lagd_description(self, lagd_toggle: bool):
@@ -118,16 +120,16 @@ class ModelsLayout(Widget):
      desc += f"<br>{tr('Actuator Delay:')} {cp:.2f} s + {tr('Software Delay:')} {sw:.2f} s = {tr('Total Delay:')} {cp + sw:.2f} s"
    self.lagd_toggle.set_description(desc)

-  def _is_downloading(self):
-    return (self.model_manager and self.model_manager.selectedBundle and
-            self.model_manager.selectedBundle.status == custom.ModelManagerSP.DownloadStatus.downloading)
-
  @staticmethod
  def calculate_cache_size():
    cache_size = 0.0
    if os.path.exists(CUSTOM_MODEL_PATH):
-      cache_size = sum(os.path.getsize(os.path.join(CUSTOM_MODEL_PATH, file)) for file in os.listdir(CUSTOM_MODEL_PATH)) / (1024**2)
-    return cache_size
+      for file in os.listdir(CUSTOM_MODEL_PATH):
+        try:
+          cache_size += os.path.getsize(os.path.join(CUSTOM_MODEL_PATH, file))
+        except OSError:
+          continue
+    return cache_size / (1024**2)

  def _clear_cache(self):
    def _callback(response):
@@ -140,36 +142,90 @@ class ModelsLayout(Widget):
    gui_app.push_widget(dialog)

  def _handle_bundle_download_progress(self):
-    self.download_item.set_visible(False)
    self.cancel_download_item.set_visible(False)
    self._downloading = False
-
-    if not self.model_manager or (not self.model_manager.selectedBundle and not self.model_manager.activeBundle):
-      return
-
-    bundle = self.model_manager.selectedBundle if self._is_downloading() or (
-      self.model_manager.selectedBundle and self.model_manager.selectedBundle.status == custom.ModelManagerSP.DownloadStatus.failed
-    ) else self.model_manager.activeBundle
-    if not bundle:
-      return
-
-    self.cancel_download_item.set_visible(bool(self.model_manager.selectedBundle) and ui_state.params.get("ModelManager_DownloadRef") is not None)
+    self._verifying = False
+    self.download_item.set_visible(True)

    if (current_time := time.monotonic()) - self.last_cache_calc_time > 0.5:
      self.last_cache_calc_time = current_time
      self.clear_cache_item.action_item.set_value(f"{self.calculate_cache_size():.2f} MB")

+    bundle = self.model_manager.selectedBundle if self.model_manager else None
+    progresses = [model.artifact.downloadProgress for model in bundle.models if model.artifact.fileName] if bundle else []
+    if not progresses or bundle.status not in (custom.ModelManagerSP.DownloadStatus.downloading,
+                                               custom.ModelManagerSP.DownloadStatus.failed):
+      self.download_item.action_item.update(name="", segments=self._slot_segments())
+      return
+
+    self.cancel_download_item.set_visible(ui_state.params.get("ModelManager_DownloadRef") is not None)
    if bundle.status == custom.ModelManagerSP.DownloadStatus.downloading:
      device._reset_interactive_timeout()

-    # every bundle is a single chunked artifact now
-    progresses = [model.artifact.downloadProgress for model in bundle.models if model.artifact.fileName]
-    if not progresses:
-      return
-
-    self.download_item.set_visible(True)
-    self.download_item.action_item.update(**self._download_row_state(progresses, bundle.internalName))
+    state = self._download_row_state(progresses, bundle.internalName)
+    if queued := queued_name(bundle.ref):
+      state["name"] += f"  |  {queued} {tr('queued')}"
+    self.download_item.action_item.update(**state)
    self._downloading = self.download_item.action_item.downloading
+    ds = custom.ModelManagerSP.DownloadStatus
+    self._verifying = any(getattr(p.status, 'raw', p.status) == ds.verifying for p in progresses)
+
+  def _slot_segments(self):
+    """small and big slots side by side; green marks the slot whose pick is actually
+    driving (runner-matched, so a failed Default big greens neither slot), an empty
+    slot shows its default."""
+    big_state = big_model_state()
+    carry_source, carry_internal, _ = carrying_model()
+    segments = []
+    for source, label in (("qcom", tr("small")), ("usbgpu", tr("big"))):
+      if segments:
+        segments.append(("|", rl.GRAY, None, None))
+      bundle = get_selected_bundle(ui_state.params, source)
+      name = bundle.internalName if bundle else default_model_name(source)
+      color = ON_COLOR if (source == carry_source and name == carry_internal) else rl.LIGHTGRAY
+      name = "● " + name
+      if source == "usbgpu":
+        if big_state == 'failed':
+          color = rl.RED
+        elif big_state == 'loading':
+          color = rl.GOLD
+      segments.append((label, rl.GRAY, None, None))
+      segments.append((name, color, None, None))
+    return segments
+
+  @staticmethod
+  def _set_item_note(item, text):
+    # a description renders only while shown; hide before clearing or the
+    # empty description keeps its visible state
+    if text:
+      item.set_description(text)
+      item.show_description(True)
+    else:
+      item.show_description(False)
+      item.set_description("")
+
+  def _status_note(self) -> str:
+    """The failover story for the Model Status row. One-way big -> small, and the
+    fallback is runner-matched: a Default big can only fall back to the Default
+    small (stock modeld), a custom big has no automatic fallback yet."""
+    if not ui_state.usbgpu:
+      return ""
+    big_bundle = get_selected_bundle(ui_state.params, "usbgpu")
+    big_name = big_bundle.internalName if big_bundle else default_model_name("usbgpu")
+    big_is_default = big_bundle is None
+    fallback_name = default_model_name("qcom")
+    state = big_model_state()
+    if state == 'failed':
+      if big_is_default:
+        return tr("Big model unavailable, {} is driving until the next drive.").format(fallback_name)
+      return tr("Big model unavailable until the next drive.")
+    if state == 'loading':
+      if big_is_default:
+        return tr("{} drives until the big model is ready.").format(fallback_name)
+      return tr("Getting the big model ready.")
+    if big_is_default:
+      return tr("{} will drive. If it fails during a drive, {} takes over until the next drive.").format(big_name, fallback_name)
+    return tr("{} will drive when the eGPU is ready.").format(big_name)

  @staticmethod
  def _download_row_state(progresses, name: str) -> dict:
@@ -182,6 +238,8 @@ class ModelsLayout(Widget):
    if ds.failed in statuses:
      # close.png is authored black and a tint cannot lift it, hence close2
      return {"name": name, "status_text": tr("download failed"), "text_color": rl.RED, "icon": "icons/close2.png"}
+    if ds.verifying in statuses:
+      return {"name": name, "downloading": True, "progress": progress, "status_text": tr("verifying")}
    if ds.downloading in statuses:
      return {"name": name, "downloading": True, "progress": progress}
    if statuses <= {ds.downloaded, ds.cached}:
@@ -203,12 +261,7 @@ class ModelsLayout(Widget):
      ui_state.params.put("ModelManager_DownloadRef", selected_bundle.ref)

  def _resolve_selected_bundle(self, ref):
-    """Finds the bundle for a ref across both hardware manifests."""
-    active = ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent)
-    source_bundles = {
-      source: self.model_manager.availableBundles if source == active else get_cached_bundles(ui_state.params, source)
-      for source in ("qcom", "usbgpu")
-    }
+    source_bundles = {source: bundles_for_source(source) for source in ("qcom", "usbgpu")}
    resolved = resolve_bundle_by_ref(ref, source_bundles)
    return resolved[0] if resolved else None

@@ -228,18 +281,10 @@ class ModelsLayout(Widget):
      folders_list.append(TreeFolder(name, [self._bundle_to_node(bundle) for bundle in folder_bundles]))

    if favorites and (fav_bundles := [bundle for bundle in bundles if bundle.ref in favorites]):
-      folders_list.insert(1, TreeFolder("Favorites", [self._bundle_to_node(bundle) for bundle in fav_bundles]))
+      folders_list.insert(0, TreeFolder("Favorites", [self._bundle_to_node(bundle) for bundle in fav_bundles]))
    return folders_list

-  def _handle_current_model_clicked(self):
-    self._open_source_dialog(ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent))
-
-  def _handle_other_model_clicked(self):
-    active = ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent)
-    self._open_source_dialog("qcom" if active == "usbgpu" else "usbgpu")
-
  def _open_source_dialog(self, source):
-    """Opens the picker for one hardware: its model folders plus the Default reset entry."""
    self._selection_source = source
    favs = ui_state.params.get("ModelManager_Favs")
    favorites = set(favs.split(';')) if favs else set()
@@ -252,12 +297,10 @@ class ModelsLayout(Widget):
    gui_app.push_widget(self.model_dialog)

  def _source_folders(self, favorites, source):
-    """Default reset entry on top, then the hardware's model folders."""
-    active = ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent)
-    bundles = self.model_manager.availableBundles if source == active else get_cached_bundles(ui_state.params, source)
+    bundles = bundles_for_source(source)
    if not bundles:
      return []
-    folders_list = [TreeFolder("", [TreeNode("Default", {'display_name': "Default"})])]
+    folders_list = [TreeFolder("", [TreeNode("Default", {'display_name': default_model_name(source)})])]
    folders_list.extend(self._get_folders(favorites, bundles))
    return folders_list

@@ -284,20 +327,27 @@ class ModelsLayout(Widget):
    self._update_lagd_description(live_delay)
    self.model_manager = ui_state.sm["modelManagerSP"]
    self._handle_bundle_download_progress()
-    source, active_name, other_name = model_info()
-    self.current_model_item.action_item.set_value(active_name)
-    self.other_model_item.set_title(tr("Big Model") if source == "qcom" else tr("Small Model"))
-    self.other_model_item.action_item.set_value(other_name)

-    if not ui_state.is_offroad():
-      self.current_model_item.action_item.set_enabled(False)
-      self.current_model_item.set_description(tr("Only available when vehicle is off, or always offroad mode is on"))
-    else:
-      self.current_model_item.action_item.set_enabled(True)
-      self.current_model_item.set_description("")
+    carry_source, _, carry_display = carrying_model()
+    for item, item_source in ((self.small_model_item, "qcom"), (self.big_model_item, "usbgpu")):
+      bundle = get_selected_bundle(ui_state.params, item_source)
+      name = bundle.displayName if bundle else default_model_name(item_source)
+      color = ON_COLOR if (item_source == carry_source and name == carry_display) else style.ITEM_TEXT_VALUE_COLOR
+      item.action_item.set_value(name, color)
+
+    note = self._status_note()
+    if note != self._last_note:
+      self._last_note = note
+      self._set_item_note(self.download_item, note)
+
+    offroad = ui_state.is_offroad()
+    self.small_model_item.action_item.set_enabled(offroad)
+    self.big_model_item.action_item.set_enabled(offroad)
+    self.small_model_item.set_description("" if offroad else tr("Only available when vehicle is off, or always offroad mode is on"))

  def _render(self, rect):
    self._scroller.render(rect)

  def show_event(self):
    self._scroller.show_event()
+    self._last_note = None  # re-expand the failover note every time the page opens
@@ -4,11 +4,14 @@ 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 math
+
 import pyray as rl
 import time
 from dataclasses import dataclass
 from openpilot.selfdrive.ui.ui_state import ui_state
 from openpilot.sunnypilot.sunnylink.api import UNREGISTERED_SUNNYLINK_DONGLE_ID
+from openpilot.system.ui.lib.application import gui_app
 from openpilot.system.ui.lib.multilang import tr_noop


@@ -18,6 +21,9 @@ METRIC_MARGIN = 30
 METRIC_START_Y = 300
 HOME_BTN = rl.Rectangle(60, 860, 180, 180)

+EGPU_ICON_WIDTH = 180
+EGPU_ICON_HEIGHT = 133
+

 # Color scheme
 class Colors:
@@ -53,6 +59,10 @@ class MetricData:
 class SidebarSP:
  def __init__(self):
    self._sunnylink_status = MetricData(tr_noop("SUNNYLINK"), tr_noop("OFFLINE"), Colors.WARNING)
+    self._egpu_green_img = gui_app.texture("icons_mici/egpu_green.png", EGPU_ICON_WIDTH, EGPU_ICON_HEIGHT)
+    self._egpu_default_img = gui_app.texture("icons_mici/egpu.png", EGPU_ICON_WIDTH, EGPU_ICON_HEIGHT)
+    self._egpu_orange_img = gui_app.texture("icons_mici/egpu_orange.png", EGPU_ICON_WIDTH, EGPU_ICON_HEIGHT)
+    self._egpu_gray_img = gui_app.texture("icons_mici/egpu_gray.png", EGPU_ICON_WIDTH, EGPU_ICON_HEIGHT)

  def _update_sunnylink_status(self):
    if not ui_state.params.get_bool("SunnylinkEnabled"):
@@ -78,6 +88,29 @@ class SidebarSP:

    self._sunnylink_status.update(tr_noop("SUNNYLINK"), status, color)

+  def _get_home_icon(self, default_img: rl.Texture) -> tuple[rl.Texture, rl.Vector2, float]:
+    default_pos = rl.Vector2(HOME_BTN.x, HOME_BTN.y)
+    if not ui_state.sm["deviceState"].chestnutPresent:
+      return default_img, default_pos, 1.0
+
+    big_model_selected = ui_state.usbgpu_compiled or ui_state.model_runner_tinygrad
+    big_model_failed = ui_state.started and ui_state.big_model_failed
+    loading = ui_state.usbgpu_loading or (big_model_selected and ui_state.started and ui_state.usbgpu_active is None)
+
+    if loading:
+      icon = self._egpu_default_img
+      opacity = 0.35 + 0.65 * (0.5 - 0.5 * math.cos(rl.get_time() * 6.0))
+    elif big_model_selected and big_model_failed:
+      icon, opacity = self._egpu_orange_img, 1.0
+    elif big_model_selected:
+      icon, opacity = self._egpu_green_img, 1.0
+    else:
+      icon, opacity = self._egpu_gray_img, 1.0
+
+    x = HOME_BTN.x + (HOME_BTN.width - icon.width) / 2
+    y = HOME_BTN.y + (HOME_BTN.height - icon.height) / 2
+    return icon, rl.Vector2(x, y), opacity
+
  def _draw_metrics_w_sunnylink(self, rect: rl.Rectangle, _temp, _panda, _connect):
    metrics = [_temp, _panda, _connect, self._sunnylink_status]
    start_y = int(rect.y) + METRIC_START_Y
@@ -4,8 +4,14 @@ 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 math
+
+import pyray as rl
+
 from openpilot.selfdrive.ui.mici.layouts.home import MiciHomeLayout
+from openpilot.selfdrive.ui.ui_state import ui_state
 from openpilot.system.ui.lib.application import FontWeight
+from openpilot.system.ui.widgets.icon_widget import IconWidget
 from openpilot.system.ui.widgets.label import UnifiedLabel


@@ -13,3 +19,35 @@ class MiciHomeLayoutSP(MiciHomeLayout):
  def __init__(self):
    super().__init__()
    self._openpilot_label = UnifiedLabel("sunnypilot", font_size=88, font_weight=FontWeight.AUDIOWIDE, max_width=480, wrap_text=False)
+    self._egpu_icon_default = IconWidget("icons_mici/egpu.png", (50, 37))
+    self._egpu_icon_default.set_visible(False)
+    self._egpu_icon_orange = IconWidget("icons_mici/egpu_orange.png", (50, 37))
+    self._egpu_icon_orange.set_visible(False)
+    gray_idx = self._status_bar_layout.widgets.index(self._egpu_icon_gray)
+    self._status_bar_layout.widgets.insert(gray_idx + 1, self._egpu_icon_default)
+    self._status_bar_layout.widgets.insert(gray_idx + 2, self._egpu_icon_orange)
+
+  def _set_egpu_visibility(self):
+    chestnut = ui_state.sm["deviceState"].chestnutPresent
+    if not chestnut:
+      self._egpu_icon.set_visible(False)
+      self._egpu_icon_default.set_visible(False)
+      self._egpu_icon_orange.set_visible(False)
+      self._egpu_icon_gray.set_visible(False)
+      return
+
+    big_model_selected = ui_state.usbgpu_compiled or ui_state.model_runner_tinygrad
+    big_model_failed = ui_state.started and ui_state.big_model_failed
+    loading = ui_state.usbgpu_loading or (big_model_selected and ui_state.started and ui_state.usbgpu_active is None)
+
+    if loading:
+      self._egpu_icon_default._opacity = 0.35 + 0.65 * (0.5 - 0.5 * math.cos(rl.get_time() * 6.0))
+      self._egpu_icon_default.set_visible(True)
+      self._egpu_icon.set_visible(False)
+      self._egpu_icon_orange.set_visible(False)
+      self._egpu_icon_gray.set_visible(False)
+    else:
+      self._egpu_icon_default.set_visible(False)
+      self._egpu_icon.set_visible(big_model_selected and not big_model_failed)
+      self._egpu_icon_orange.set_visible(big_model_selected and big_model_failed)
+      self._egpu_icon_gray.set_visible(not big_model_selected)
@@ -8,11 +8,11 @@ import pyray as rl

 from openpilot.cereal import custom
 from openpilot.selfdrive.ui.mici.widgets.dialog import BigDialog
-from openpilot.sunnypilot.models.fetcher import ModelFetcher, get_cached_bundles
-from openpilot.sunnypilot.models.helpers import ACTIVE_BUNDLE_KEYS
+from openpilot.sunnypilot.models.helpers import ACTIVE_BUNDLE_KEYS, get_selected_bundle
 from openpilot.selfdrive.ui.mici.widgets.button import BigButton
 from openpilot.selfdrive.ui.ui_state import ui_state, device
-from openpilot.selfdrive.ui.sunnypilot.model_info import model_info
+from openpilot.selfdrive.ui.sunnypilot.model_info import (active_source, big_model_state, bundles_for_source, carrying_model,
+                                                           default_model_name, model_info, queued_name)
 from openpilot.system.ui.lib.application import FontWeight, gui_app
 from openpilot.system.ui.lib.multilang import tr
 from openpilot.system.ui.widgets import Widget
@@ -20,10 +20,22 @@ from openpilot.system.ui.widgets.label import UnifiedLabel
 from openpilot.system.ui.widgets.scroller import NavScroller

 def _model_info() -> tuple[str, str, str]:
-  """(active model, other-model header, other-model text) for the panel."""
+  """(active model, info header, info text) for the panel. Runner-matched: the
+  active line names what actually drives, and a notable big-model state takes
+  the info pair."""
  source, active_name, other_name = model_info()
+  state = big_model_state()
+  _, _, carry_display = carrying_model()
+  if carry_display is None:
+    big = get_selected_bundle(ui_state.params, "usbgpu")
+    carry_display = big.displayName if big else default_model_name("usbgpu")
+  active_text = (carry_display or active_name).lower()
+  if state == 'failed':
+    return active_text, tr("big model"), tr("unavailable")
+  if state == 'loading':
+    return active_text, tr("big model"), tr("getting ready")
  header = tr("small model") if source == "usbgpu" else tr("big model")
-  return active_name.lower(), header, other_name.lower()
+  return active_text, header, other_name.lower()


 class CurrentModelInfo(Widget):
@@ -100,8 +112,13 @@ class ModelsLayoutMici(NavScroller):
    self.focused_widget = self.select_model_btn

    hardware_btns = []
+    active = active_source()
    for source, label in (("qcom", tr("small models")), ("usbgpu", tr("big models"))):
-      btn = BigButton(label.lower())
+      bundle = get_selected_bundle(ui_state.params, source)
+      value = (bundle.internalName if bundle else default_model_name(source)).lower()
+      if source == active:
+        value += f" ({tr('active')})"
+      btn = BigButton(label.lower(), value=value)
      btn.set_click_callback(lambda s=source: self._select_hardware(s))
      hardware_btns.append(btn)
    self._push_selection_view(hardware_btns)
@@ -111,20 +128,16 @@ class ModelsLayoutMici(NavScroller):

    favs = ui_state.params.get("ModelManager_Favs")
    favorites = set(favs.split(';')) if favs else set()
-    active = ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent)

-    if source != active:
-      bundles = get_cached_bundles(ui_state.params, source)
-      if not bundles:
-        gui_app.push_widget(BigDialog(title=tr("No models available"),
-                                      description=tr("No models are available for this hardware yet. Connect to the internet and refresh the model list.")))
-        return
-    else:
-      bundles = self.model_manager.availableBundles
+    bundles = bundles_for_source(source)
+    if not bundles:
+      gui_app.push_widget(BigDialog(title=tr("No models available"),
+                                    description=tr("No models are available for this hardware yet. Connect to the internet and refresh the model list.")))
+      return
    folders = self._get_grouped_bundles(bundles, favorites)

    folder_buttons = []
-    default_btn = BigButton(tr("default"))
+    default_btn = BigButton(default_model_name(source).lower())
    default_btn.set_click_callback(lambda s=source: self._select_default(s))
    folder_buttons.append(default_btn)

@@ -155,13 +168,8 @@ class ModelsLayoutMici(NavScroller):
      return
    favs = ui_state.params.get("ModelManager_Favs")
    favorites = set(favs.split(';')) if favs else set()
-    active = ModelFetcher.active_source(ui_state.sm["deviceState"].chestnutPresent)

-    if source != active:
-      bundles = get_cached_bundles(ui_state.params, source)
-    else:
-      bundles = self.model_manager.availableBundles
-    folders = self._get_grouped_bundles(bundles, favorites)
+    folders = self._get_grouped_bundles(bundles_for_source(source), favorites)
    bundles = sorted(folders.get(folder_name, []), key=lambda b: b.index, reverse=True)

    btns = []
@@ -212,18 +220,25 @@ class ModelsLayoutMici(NavScroller):
      device.set_override_interactive_timeout(5)
      progress = 0.0
      count = 0
+      verifying = False
      for model in manager.selectedBundle.models:
        count += 1
        p = model.artifact.downloadProgress
-        if p.status == custom.ModelManagerSP.DownloadStatus.downloading:
+        if p.status in (custom.ModelManagerSP.DownloadStatus.downloading,
+                        custom.ModelManagerSP.DownloadStatus.verifying):
          progress += p.progress
+          verifying = verifying or p.status == custom.ModelManagerSP.DownloadStatus.verifying
        elif p.status in (custom.ModelManagerSP.DownloadStatus.downloaded,
                          custom.ModelManagerSP.DownloadStatus.cached):
          progress += 100.0

-      self.current_model_info.current_model_header.set_text(tr("downloading"))
+      self.current_model_info.current_model_header.set_text(tr("verifying") if verifying else tr("downloading"))
+      self.cancel_download_btn.set_text(tr("cancel verification") if verifying else tr("cancel download"))
      self.current_model_info.current_model_header._shimmer = True
-      self.current_model_info.current_model_text.set_text(f"{manager.selectedBundle.internalName.lower()}")
+      name_text = manager.selectedBundle.internalName.lower()
+      if queued := queued_name(manager.selectedBundle.ref):
+        name_text += f"  |  {queued.lower()} {tr('queued')}"
+      self.current_model_info.current_model_text.set_text(name_text)
      self.current_model_info.info_header.set_text(tr("progress") + self._download_progress)
      self.current_model_info.info_header._shimmer = True
      self.current_model_info.info_text.set_text(f"{progress/count:.2f}%")
@@ -7,6 +7,7 @@ See the LICENSE.md file in the root directory for more details.
 import pyray as rl

 from openpilot.selfdrive.ui.mici.onroad.hud_renderer import HudRenderer
+from openpilot.selfdrive.ui.ui_state import ui_state
 from openpilot.selfdrive.ui.sunnypilot.onroad.blind_spot_indicators import BlindSpotIndicators


@@ -21,6 +22,8 @@ class HudRendererSP(HudRenderer):

  def _render(self, rect: rl.Rectangle) -> None:
    super()._render(rect)
+    if ui_state.usbgpu and not ui_state.usbgpu_compiled and ui_state.model_runner_tinygrad:
+      self._draw_model_source(rect)
    self.blind_spot_indicators.render(rect)

  def _has_blind_spot_detected(self) -> bool:
@@ -5,21 +5,84 @@ 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 openpilot.selfdrive.ui.ui_state import ui_state
-from openpilot.sunnypilot.models.helpers import get_active_source, get_selected_bundle
+from openpilot.sunnypilot.models.fetcher import get_cached_bundles
+from openpilot.sunnypilot.models.helpers import get_active_source, get_selected_bundle, resolve_bundle_by_ref
 from openpilot.sunnypilot.models.model_name import DEFAULT_BIG_MODEL, DEFAULT_MODEL


+def active_source() -> str:
+  return get_active_source(usbgpu=ui_state.usbgpu,
+                           usbgpu_active=ui_state.usbgpu_active, usbgpu_loading=ui_state.usbgpu_loading,
+                           offroad=ui_state.is_offroad())
+
+
+def bundles_for_source(source: str):
+  if source == active_source():
+    return ui_state.sm["modelManagerSP"].availableBundles
+  return get_cached_bundles(ui_state.params, source)
+
+
+def default_model(source: str) -> str:
+  return DEFAULT_BIG_MODEL if source == 'usbgpu' else DEFAULT_MODEL
+
+
+def default_model_name(source: str) -> str:
+  return f"{default_model(source)} (Default)"
+
+
+def big_model_state() -> str | None:
+  """'failed' | 'loading' | None, mirroring the sidebar's detection (#1969)."""
+  if ui_state.started and ui_state.usbgpu and ui_state.big_model_failed:
+    return 'failed'
+  big_selected = ui_state.usbgpu_compiled or ui_state.model_runner_tinygrad
+  if ui_state.usbgpu_loading or (big_selected and ui_state.started and ui_state.usbgpu_active is None):
+    return 'loading'
+  return None
+
+
+def carrying_model() -> tuple[str | None, str | None, str | None]:
+  """(source, internal name, display name) of what actually drives. Runner-matched:
+  when a Default big cannot carry, stock modeld runs the Default small, never the
+  small slot's pick; a custom big has no automatic fallback yet -> (None, None, None)."""
+  source = active_source()
+  if source == "usbgpu":
+    bundle = get_selected_bundle(ui_state.params, "usbgpu")
+    if bundle:
+      return "usbgpu", bundle.internalName, bundle.displayName
+    name = default_model_name("usbgpu")
+    return "usbgpu", name, name
+  if ui_state.usbgpu:
+    if get_selected_bundle(ui_state.params, "usbgpu") is None:
+      name = default_model_name("qcom")
+      return "qcom", name, name
+    return None, None, None
+  bundle = get_selected_bundle(ui_state.params, "qcom")
+  if bundle:
+    return "qcom", bundle.internalName, bundle.displayName
+  name = default_model_name("qcom")
+  return "qcom", name, name
+
+
+def queued_name(current_ref) -> str | None:
+  ref = ui_state.params.get("ModelManager_DownloadRef")
+  if ref and ref != current_ref:
+    source_bundles = {source: bundles_for_source(source) for source in ("qcom", "usbgpu")}
+    if resolved := resolve_bundle_by_ref(ref, source_bundles):
+      return resolved[0].internalName
+  return None
+
+
 def model_info() -> tuple[str, str, str]:
-  """returns (active source, active model name, other model name)"""
-  source = get_active_source(usbgpu=ui_state.usbgpu,
-                             usbgpu_active=ui_state.usbgpu_active, usbgpu_loading=ui_state.usbgpu_loading,
-                             offroad=ui_state.is_offroad())
+  """returns (active source, active model name, other model name)
+
+  Names come from the params slots, never modelManagerSP.activeBundle — the
+  manager republishes a tick after a chestnut change, so the stale bundle
+  would flash the wrong model."""
+  source = active_source()
  other = "qcom" if source == "usbgpu" else "usbgpu"
  active_bundle = get_selected_bundle(ui_state.params, source)
  other_bundle = get_selected_bundle(ui_state.params, other)

-  active_name = active_bundle.displayName if active_bundle \
-    else f"{DEFAULT_BIG_MODEL if source == 'usbgpu' else DEFAULT_MODEL} (Default)"
-  other_name = other_bundle.displayName if other_bundle \
-    else f"{DEFAULT_MODEL if source == 'usbgpu' else DEFAULT_BIG_MODEL} (Default)"
+  active_name = active_bundle.displayName if active_bundle else default_model_name(source)
+  other_name = other_bundle.displayName if other_bundle else default_model_name(other)
  return source, active_name, other_name
@@ -10,7 +10,7 @@ from openpilot.cereal import messaging, log, custom
 from opendbc.car.structs import car
 from openpilot.common.params import Params
 from openpilot.selfdrive.ui.sunnypilot.layouts.settings.display import OnroadBrightness
-from openpilot.sunnypilot.models.helpers import get_active_bundle
+from openpilot.sunnypilot.models.helpers import ACTIVE_BUNDLE_KEYS, get_active_source
 from openpilot.sunnypilot.sunnylink.sunnylink_state import SunnylinkState
 from openpilot.system.ui.lib.application import gui_app
 from openpilot.system.ui.sunnypilot.widgets.screen_saver import ScreenSaverSP
@@ -44,6 +44,7 @@ class UIStateSP:
    self.screensaver_enabled: bool = False

    self.active_bundle = None
+    self.model_runner_tinygrad: bool = False
    self.blindspot: bool = False
    self.chevron_metrics = None
    self.custom_interactive_timeout: int = 0
@@ -151,7 +152,10 @@ class UIStateSP:
      self.has_icbm = self.CP_SP.intelligentCruiseButtonManagementAvailable and self.params.get_bool("IntelligentCruiseButtonManagement")

    self._enforce_constraints()
-    self.active_bundle = get_active_bundle(self.params)
+    source = get_active_source(usbgpu=self.usbgpu, usbgpu_active=self.usbgpu_active,
+                               usbgpu_loading=self.usbgpu_loading, offroad=self.is_offroad())
+    self.active_bundle = self.params.get(ACTIVE_BUNDLE_KEYS[source])
+    self.model_runner_tinygrad = self.active_bundle is not None and self.active_bundle.get("runner") == "tinygrad"
    self.blindspot = self.params.get_bool("BlindSpot")
    self.chevron_metrics = self.params.get("ChevronInfo")
    self.custom_interactive_timeout = self.params.get("InteractivityTimeout", return_default=True)
@@ -112,6 +112,15 @@ class UIState(UIStateSP):
  def add_on_body_changed_callbacks(self, callback: Callable[[], None]):
    self._on_body_changed_callbacks.append(callback)

+  @property
+  def big_model_failed(self) -> bool:
+    # Mirrors the onroad HUD's four-condition check so sidebar and home icons reflect the same failure states
+    return (self.usbgpu_active is False or
+            not self.sm['deviceState'].chestnutPresent or
+            (self.usbgpu_active is True and self.sm.recv_frame['modelV2'] > self.started_frame and
+             not self.sm.alive['modelV2']) or
+            (self.usbgpu_active is None and self.sm.recv_frame['modelV2'] > self.started_frame))
+
  @property
  def engaged(self) -> bool:
    return self.started and (self.sm["selfdriveState"].enabled or self.sm["selfdriveStateSP"].mads.enabled)
@@ -7,6 +7,7 @@ See the LICENSE.md file in the root directory for more details.
 """

 import argparse
+import math
 import os
 import tempfile
 import time
@@ -66,14 +67,15 @@ def get_policy_npy_shapes(input_shapes: dict, is_supercombo: bool = False) -> tu
  if desire_key:
    shapes['desire'] = (input_shapes[desire_key][2],)

-  if is_supercombo and 'features_buffer' in input_shapes:
-    fb = input_shapes['features_buffer']
-    shapes['prev_feat'] = (fb[0], fb[2])
-
  for key, shape in input_shapes.items():
    if key not in (desire_key, 'features_buffer') and 'img' not in key:
      shapes[key] = tuple(shape)

+  if is_supercombo and 'features_buffer' in input_shapes:
+    fb = input_shapes['features_buffer']
+    feat_dim = math.prod(fb[2:])
+    shapes['prev_feat'] = (fb[0], feat_dim)
+
  sizes = [int(np.prod(size)) for size in shapes.values()]
  return shapes, sizes

@@ -117,8 +119,9 @@ def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = D
  }

  if features_buffer:
+    feat_dim = math.prod(features_buffer[2:])
    feat_q_len = frame_skip * features_buffer[1] if is_supercombo else frame_skip * (features_buffer[1] - 1) + 1
-    queues['feat_q'] = Tensor(np.zeros((feat_q_len, features_buffer[0], features_buffer[2]),
+    queues['feat_q'] = Tensor(np.zeros((feat_q_len, features_buffer[0], feat_dim),
                       dtype=np.float32), device=device).contiguous().realize()

  queues.update({key: Tensor(value, device='NPY').realize() for key, value in npy_arrays.items() if key in ('tfm', 'big_tfm')})
@@ -183,14 +186,14 @@ 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).realize()
-    big_img = shift_and_sample(big_img_q, warped_dev[1:2], sample_skip_fn).realize()
+    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)

    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).realize()
+    desire_buf = 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():
@@ -199,7 +202,7 @@ 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).realize()
+      inputs['features_buffer'] = 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()
@@ -211,7 +214,7 @@ def make_run_policy(vision_runner, policy_runners: list, features_slice: slice,

    inputs.update({road_key: img, wide_key: big_img})
    if 'features_buffer' not in inputs:
-      inputs['features_buffer'] = sample_skip_fn(feat_q)
+      inputs['features_buffer'] = sample_skip_fn(feat_q).reshape(input_shapes['features_buffer'])

    policy_out = next(iter(policy_runners[0](inputs).values())).cast('float32').realize()
    if 'features_buffer' not in inputs and features_slice is not None:
@@ -91,7 +91,7 @@ class ModelState(ModelStateBase):
    if env_pkl and os.path.exists(env_pkl):
      model_bundle = None
    else:
-      model_bundle = get_active_bundle()
+      model_bundle = get_active_bundle(usbgpu=usbgpu)
    self.generation = model_bundle.generation if model_bundle is not None else None
    overrides = {override.key: override.value for override in model_bundle.overrides} if model_bundle else {}

@@ -190,8 +190,8 @@ def tmp_path():

 def patch_modeld(monkeypatch):
  def _patch(bundle):
-    monkeypatch.setattr(helpers, 'get_active_bundle', lambda params=None: bundle)
-    monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None: bundle)
+    monkeypatch.setattr(helpers, 'get_active_bundle', lambda params=None, *, usbgpu=None: bundle)
+    monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None, *, usbgpu=None: bundle)

  return _patch

@@ -59,8 +59,8 @@ class TestFindDrivingPkl(OpenpilotTestCase):
 class TestModelStateCombinedInit(OpenpilotTestCase):
  def test_asserts_when_no_pkl(self, monkeypatch):
    bundle = DummyBundle(models=[], is_20hz=True)
-    monkeypatch.setattr(helpers, 'get_active_bundle', lambda params=None: bundle)
-    monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None: bundle)
+    monkeypatch.setattr(helpers, 'get_active_bundle', lambda params=None, *, usbgpu=None: bundle)
+    monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None, *, usbgpu=None: bundle)
    with self.assertRaisesRegex(AssertionError, "No driving pkl found"):
      ModelState(cam_w=CAM_W, cam_h=CAM_H)

@@ -195,3 +195,85 @@ class TestReadFileChunkedToDisk(OpenpilotTestCase):

      assert out.parent == Path(d)
      assert out.read_bytes() == payload
+
+
+class Test4DFeaturesBuffer(OpenpilotTestCase):
+  def test_get_policy_npy_shapes_4d(self):
+    from openpilot.sunnypilot.modeld_v2.compile_modeld import get_policy_npy_shapes
+    input_shapes = {
+      'desire_pulse': (1, 25, 8),
+      'features_buffer': (1, 24, 32, 512),  # compare 4d to 3d for regression
+      'traffic_convention': (1, 2),
+      'action_t': (1, 2)
+    }
+    shapes, sizes = get_policy_npy_shapes(input_shapes, is_supercombo=True)
+    assert shapes['prev_feat'] == (1, 16384)
+    assert sizes == [8, 2, 2, 16384]
+
+  def test_get_policy_npy_shapes_3d(self):
+    from openpilot.sunnypilot.modeld_v2.compile_modeld import get_policy_npy_shapes
+    input_shapes = {
+      'desire_pulse': (1, 25, 8),
+      'features_buffer': (1, 24, 512),
+      'traffic_convention': (1, 2),
+      'action_t': (1, 2)
+    }
+    shapes, sizes = get_policy_npy_shapes(input_shapes, is_supercombo=True)
+    assert shapes['prev_feat'] == (1, 512)
+    assert sizes == [8, 2, 2, 512]
+
+
+class TestStockCompileModeldEquivalence(OpenpilotTestCase):
+  def test_get_policy_npy_shapes_matches_stock(self):
+    from openpilot.selfdrive.modeld.compile_modeld import get_policy_npy_shapes as stock_get_policy_npy_shapes
+    from openpilot.sunnypilot.modeld_v2.compile_modeld import get_policy_npy_shapes as sunny_get_policy_npy_shapes
+
+    stock_input_shapes = {
+      'desire_pulse': (1, 25, 8),
+      'features_buffer': (1, 24, 512), # see below comment
+      'traffic_convention': (1, 2),
+      'action_t': (1, 2),
+    }
+
+    stock_shapes, stock_sizes = stock_get_policy_npy_shapes(stock_input_shapes)
+    sunny_shapes, sunny_sizes = sunny_get_policy_npy_shapes(stock_input_shapes, is_supercombo=True)
+
+    assert sunny_shapes == stock_shapes
+    assert sunny_sizes == stock_sizes
+    assert sunny_shapes['prev_feat'] == (1, 512)
+
+  def test_make_input_queues_full_stock_equivalence(self):
+    from openpilot.selfdrive.modeld.compile_modeld import make_input_queues as stock_make_input_queues
+    from openpilot.sunnypilot.modeld_v2.compile_modeld import make_supercombo_input_queues as sunny_make_supercombo_input_queues
+    input_shapes = {
+      'img': (1, 12, 128, 256),
+      'desire_pulse': (1, 25, 8),
+      'features_buffer': (1, 24, 512),  # when https://github.com/commaai/openpilot/pull/38681 merges, update to 1,24,32,512
+      'traffic_convention': (1, 2),
+      'action_t': (1, 2),
+    }
+    frame_skip = 4
+
+    stock_queues, stock_npy = stock_make_input_queues(input_shapes, frame_skip, device='NPY')
+    sunny_queues, sunny_npy = sunny_make_supercombo_input_queues(input_shapes, frame_skip, device='NPY')
+    assert set(sunny_queues.keys()) == set(stock_queues.keys())
+    for key in stock_queues:
+      assert sunny_queues[key].shape == stock_queues[key].shape, \
+        f"Queue shape mismatch for {key}: sunny {sunny_queues[key].shape} != stock {stock_queues[key].shape}"
+    assert set(sunny_npy.keys()) == set(stock_npy.keys())
+    for key in stock_npy:
+      assert sunny_npy[key].shape == stock_npy[key].shape, \
+        f"Numpy array shape mismatch for {key}: sunny {sunny_npy[key].shape} != stock {stock_npy[key].shape}"
+
+  def test_make_warp_queues_stock_equivalence(self):
+    from openpilot.selfdrive.modeld.compile_modeld import make_warp_input_queues as stock_make_warp_queues
+    from openpilot.sunnypilot.modeld_v2.compile_modeld import make_warp_queues as sunny_make_warp_queues
+    stock_vision_shapes = {'img': (1, 12, 128, 256)}  # for now?
+    stock_queues, stock_npy = stock_make_warp_queues(stock_vision_shapes, frame_skip=4, device='NPY')
+    sunny_queues, sunny_npy = sunny_make_warp_queues(device='NPY')
+
+    assert set(sunny_npy.keys()) == set(stock_npy.keys()) == {'tfm', 'big_tfm'}
+    for key in sunny_npy:
+      assert sunny_npy[key].shape == stock_npy[key].shape == (3, 3)
+
+
@@ -138,8 +138,8 @@ 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_v20.json"
-  MODEL_URL_USBGPU = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_usbgpu_v21.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v21.json"
+  MODEL_URL_USBGPU = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_usbgpu_v22.json"

  MODEL_SOURCES = {
    "qcom": (MODEL_URL, ""),
@@ -149,27 +149,20 @@ class ModelFetcher:
  def __init__(self, params: Params):
    self.params = params
    self.model_parser = ModelParser()
-    self._active_json_published = False
    self.model_caches = {
      source: ModelCache(params, suffix=suffix)
      for source, (_, suffix) in self.MODEL_SOURCES.items()
    }
-    self.model_url = self.MODEL_URL
-    self._update_model_source()
+    self._refetched: set[str] = set()
+    self.params.put("ModelManager_ActiveJson", {
+      "qcom": self.MODEL_URL,
+      "usbgpu": self.MODEL_URL_USBGPU,
+    }, block=True)

  @staticmethod
  def active_source(chestnut_present: bool) -> str:
    return "usbgpu" if chestnut_present else "qcom"

-  def _update_model_source(self) -> None:
-    """Publishes the manifest URLs for both sources"""
-    if not self._active_json_published:
-      self._active_json_published = True
-      self.params.put("ModelManager_ActiveJson", {
-        "qcom": self.MODEL_URL,
-        "usbgpu": self.MODEL_URL_USBGPU,
-      }, block=True)
-
  def _fetch_and_cache_models(self, source: str) -> list[custom.ModelManagerSP.ModelBundle] | None:
    """Fetches fresh model data from remote and updates cache.
    Returns None on transport errors. Raises on 404 and other fatal HTTP errors.
@@ -206,17 +199,22 @@ class ModelFetcher:

  @staticmethod
  def _cache_matches_source(source: str, cached_data: dict) -> bool:
-    """Confirms a cached manifest contains requested source's models."""
    bundles = cached_data.get("bundles", [])
    if source == "usbgpu":
      return any(bundle.get("is_big") is True for bundle in bundles)
    return not any(bundle.get("is_big") is True for bundle in bundles)

-  def _get_source_bundles(self, source: str) -> list[custom.ModelManagerSP.ModelBundle]:
+  def get_bundles_for_source(self, source: str) -> list[custom.ModelManagerSP.ModelBundle]:
+    if source not in self.MODEL_SOURCES:
+      cloudlog.warning(f"Unknown model source: {source}")
+      return []
+
    cached_data, is_expired = self.model_caches[source].get()

    if cached_data and not is_expired:
-      if self._cache_matches_source(source, cached_data):
+      # a source is refetched over a mismatch at most once per process: if the fresh
+      # manifest still mismatches, the URL is authoritative and the cache is trusted
+      if self._cache_matches_source(source, cached_data) or source in self._refetched:
        try:
          parsed = self.model_parser.parse_models(cached_data)
        except Exception:
@@ -229,7 +227,8 @@ class ModelFetcher:
          # manifest version) - do not trust it, refetch so the source is repopulated
          cloudlog.warning(f"Cached models for {source} have no valid bundles; refetching")
      else:
-        cloudlog.warning(f"Cached models for {source} not valid; refetching")
+        self._refetched.add(source)
+        cloudlog.warning(f"Cached models for {source} not valid; refetching once")

    fetched_bundles = self._fetch_and_cache_models(source)
    if fetched_bundles is not None:
@@ -244,16 +243,8 @@ class ModelFetcher:
    except Exception:
      return []

-  def get_bundles_for_source(self, source: str) -> list[custom.ModelManagerSP.ModelBundle]:
-    """Gets the list of available models for a specific source, with smart cache handling."""
-    if source not in self.MODEL_SOURCES:
-      cloudlog.warning(f"Unknown model source: {source}")
-      return []
-    return self._get_source_bundles(source)
-

 def get_cached_bundles(params: Params, source: str) -> list[custom.ModelManagerSP.ModelBundle]:
-  """Reads a source's cached manifest from params and parses it into bundles."""

  if source not in ModelFetcher.MODEL_SOURCES:
    cloudlog.warning(f"Unknown model source: {source}")
@@ -19,7 +19,7 @@ from openpilot.common.hardware.hw import Paths
 from openpilot.selfdrive.modeld.helpers import usbgpu_present

 # SET ME TO THE EXACT JSON VERSION WE SET IN SUNNYPILOT_MODELS REPO
-REQUIRED_JSON_VERSION = 17
+REQUIRED_JSON_VERSION = 18

 CUSTOM_MODEL_PATH = Paths.model_root()
 METADATA_PATH = Path(__file__).parent / '../models/supercombo_metadata.pkl'
@@ -29,7 +29,7 @@ ACTIVE_BUNDLE_KEYS = {
  "qcom": "ModelManager_ActiveBundle",
  "usbgpu": "ModelManager_ActiveBundleUSBGPU",
 }
-_LAST_VALIDATED_RAW: dict[str, bytes | None] = {}
+_LAST_VALIDATED_RAW: dict[str, dict | None] = {}


 def _compute_hash(file_path: str) -> str | None:
@@ -92,11 +92,11 @@ def _bundle_needs_reset(active_bundle: custom.ModelManagerSP.ModelBundle, availa
  if available_bundles is not None:
    matching_bundle = None
    for bundle in available_bundles:
-      if getattr(active_bundle, 'ref', None) and getattr(bundle, 'ref', None):
+      if active_bundle.ref and bundle.ref:
        if active_bundle.ref == bundle.ref:
          matching_bundle = bundle
          break
-      elif getattr(active_bundle, 'internalName', None) == getattr(bundle, 'internalName', None):
+      elif active_bundle.internalName == bundle.internalName:
        matching_bundle = bundle
        break

@@ -104,12 +104,8 @@ def _bundle_needs_reset(active_bundle: custom.ModelManagerSP.ModelBundle, availa
      return True
    if active_bundle.minimumSelectorVersion != matching_bundle.minimumSelectorVersion:
      return True
-
-    active_runner = getattr(active_bundle, 'runner', None)
-    matching_runner = getattr(matching_bundle, 'runner', None)
-    if active_runner is not None and matching_runner is not None:
-      if getattr(active_runner, 'raw', active_runner) != getattr(matching_runner, 'raw', matching_runner):
-        return True
+    if active_bundle.runner != matching_bundle.runner:
+      return True
    if set(_bundle_artifacts(active_bundle)) != set(_bundle_artifacts(matching_bundle)):
      return True

@@ -127,7 +123,7 @@ def _parse_active_bundle(raw_bundle) -> "custom.ModelManagerSP.ModelBundle | Non

 def get_selected_bundle(params: Params | None = None, source: str = "qcom") -> "custom.ModelManagerSP.ModelBundle | None":
  params = params or Params()
-  return _parse_active_bundle(params.get(ACTIVE_BUNDLE_KEYS.get(source, "ModelManager_ActiveBundle")))
+  return _parse_active_bundle(params.get(ACTIVE_BUNDLE_KEYS[source]))


 def get_active_source(usbgpu: bool | None = None, usbgpu_active: bool | None = None,
@@ -140,17 +136,15 @@ def get_active_source(usbgpu: bool | None = None, usbgpu_active: bool | None = N


 def get_active_bundle(params: Params | None = None, *, usbgpu: bool | None = None) -> "custom.ModelManagerSP.ModelBundle | None":
+  # no cross-slot fallback: an empty active slot means the hardware default, which
+  # only stock modeld can run - modeld_v2 requires a real bundle
  params = params or Params()
-  if get_active_source(usbgpu=usbgpu) == "usbgpu":
-    if bundle := get_selected_bundle(params, "usbgpu"):
-      return bundle
-  return get_selected_bundle(params, "qcom")
+  return get_selected_bundle(params, get_active_source(usbgpu=usbgpu))


 def resolve_bundle_by_ref(
  ref: str, source_bundles: dict[str, list[custom.ModelManagerSP.ModelBundle]],
 ) -> "tuple[custom.ModelManagerSP.ModelBundle, str] | None":
-  """Finds the bundle matching a ref across all sources."""
  for source, bundles in source_bundles.items():
    for bundle in bundles:
      if bundle.ref == ref:
@@ -173,15 +167,16 @@ def _validate_active_bundle(params: Params, source: str, available_bundles: list
  if active_bundle is None or _bundle_needs_reset(active_bundle, available_bundles):
    cloudlog.warning(f"Active model bundle invalid for {source}; resetting to default")
    params.remove(key)
-    params.put("ModelRunnerTypeCache", int(custom.ModelManagerSP.Runner.stock), block=True)
    _LAST_VALIDATED_RAW[key] = None
  else:
    _LAST_VALIDATED_RAW[key] = raw_bundle


 def validate_active_bundles(params: Params, source_bundles: dict[str, list[custom.ModelManagerSP.ModelBundle]]) -> None:
+  # an empty list means the fetch failed, not that the catalog dropped the bundle
  for source, bundles in source_bundles.items():
-    _validate_active_bundle(params, source, bundles)
+    _validate_active_bundle(params, source, bundles or None)
+  get_active_model_runner(params, force_check=True)


 def get_active_model_runner(params: Params | None = None, force_check: bool = False) -> int:
@@ -24,6 +24,10 @@ from openpilot.sunnypilot.models.helpers import (ACTIVE_BUNDLE_KEYS, get_active_
 DOWNLOAD_TIMEOUT = (30, 30)


+class DownloadCancelled(Exception):
+  pass
+
+
 class ModelManagerSP:
  """Manages model downloads and status reporting"""

@@ -39,6 +43,17 @@ class ModelManagerSP:
    self.active_bundle: custom.ModelManagerSP.ModelBundle = get_active_bundle(self.params, usbgpu=self.chestnut_present)
    self._chunk_size = 128 * 1000  # 128 KB chunks
    self._download_start_times: dict[str, float] = {}  # Track start time per model
+    self._download_ref: bytes | str | None = None
+
+  def _download_interrupted(self) -> bool:
+    # only removal cancels: a different ref is a queued selection that
+    # _release_download_ref leaves in place for the next tick
+    return self.params.get("ModelManager_DownloadRef") is None
+
+  def _release_download_ref(self) -> None:
+    if self.params.get("ModelManager_DownloadRef") == self._download_ref:
+      self.params.remove("ModelManager_DownloadRef")
+    self._download_ref = None

  def _sync_artifact_progress(self, source_artifact) -> None:
    """Mirror download progress to all artifacts sharing the same filename in the selected bundle."""
@@ -80,8 +95,8 @@ class ModelManagerSP:
          f.write(chunk)
          bytes_downloaded += len(chunk)

-          if self.params.get("ModelManager_DownloadRef") is None:
-            raise Exception("Download cancelled")
+          if self._download_interrupted():
+            raise DownloadCancelled("Download cancelled")

          if total_size > 0:
            progress = (bytes_downloaded / total_size) * 100
@@ -94,7 +109,7 @@ class ModelManagerSP:
    # Clean up start time after download completes
    del self._download_start_times[model.fileName]

-  async def _download_chunked(self, base_url: str, base_path: str, artifact) -> None:
+  async def _download_chunked(self, base_url: str, base_path: str, artifact, skip: frozenset[int] | set[int] = frozenset()) -> None:
    from openpilot.common.file_chunker import get_chunk_name, get_manifest_path

    num_chunks = len(artifact.chunks)
@@ -106,8 +121,11 @@ class ModelManagerSP:

    # Shared connection saves a TCP+TLS handshake per chunk.
    # Keep sequential: the link saturates on one stream and Session is not thread-safe.
+    completed = len(skip)
    with requests.Session() as session:
      for i, _ in enumerate(artifact.chunks):
+        if i in skip:
+          continue
        chunk_url = get_chunk_name(base_url, i, num_chunks)
        chunk_path = get_chunk_name(base_path, i, num_chunks)
        chunk_downloaded = 0
@@ -118,15 +136,16 @@ class ModelManagerSP:
            for data in response.iter_content(chunk_size=self._chunk_size):
              f.write(data)
              chunk_downloaded += len(data)
-              if self.params.get("ModelManager_DownloadRef") is None:
-                raise Exception("Download cancelled")
+              if self._download_interrupted():
+                raise DownloadCancelled("Download cancelled")
              intra = chunk_downloaded / max(chunk_size, 1)
-              progress = min(99.0, ((i + intra) / num_chunks) * 100)
+              progress = min(99.0, ((completed + intra) / num_chunks) * 100)
              artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloading
              artifact.downloadProgress.progress = progress
              artifact.downloadProgress.eta = self._calculate_eta(artifact.fileName, progress)
              self._sync_artifact_progress(artifact)
              self._report_status()
+        completed += 1

    with open(manifest_path, 'w') as f:  # noqa: ASYNC230
      f.write(str(num_chunks))
@@ -137,6 +156,8 @@ class ModelManagerSP:
  async def _process_artifact(self, artifact, destination_path: str) -> None:
    if not artifact.downloadUri.uri:
      return None
+    if self._download_interrupted():
+      raise DownloadCancelled("Download cancelled")

    url = artifact.downloadUri.uri
    expected_hash = artifact.downloadUri.sha256
@@ -144,21 +165,23 @@ class ModelManagerSP:
    full_path = os.path.join(destination_path, filename)

    try:
+      # progress counts only valid chunks so a resumed download continues the
+      # bar from where verification left it, instead of falling back to zero
      is_cached = False
+      valid_chunks: set[int] = set()
      if len(artifact.chunks) > 0:
        from openpilot.common.file_chunker import get_chunk_name
        num_chunks = len(artifact.chunks)
-        chunks_valid = True
        for i, chunk in enumerate(artifact.chunks):
-          chunk_path = get_chunk_name(full_path, i, num_chunks)
-          if not await verify_file(chunk_path, chunk.sha256):
-            chunks_valid = False
-            break
-          artifact.downloadProgress.progress = ((i + 1) / num_chunks) * 100
+          if self._download_interrupted():
+            raise DownloadCancelled("Download cancelled")
+          if await verify_file(get_chunk_name(full_path, i, num_chunks), chunk.sha256):
+            valid_chunks.add(i)
+          artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.verifying
+          artifact.downloadProgress.progress = (len(valid_chunks) / num_chunks) * 100
          self._sync_artifact_progress(artifact)
          self._report_status()
-        if chunks_valid and num_chunks > 0:
-          is_cached = True
+        is_cached = len(valid_chunks) == num_chunks
      else:
        if await verify_file(full_path, expected_hash):
          is_cached = True
@@ -172,7 +195,7 @@ class ModelManagerSP:
        return

      if len(artifact.chunks) > 0:
-        await self._download_chunked(url, full_path, artifact)
+        await self._download_chunked(url, full_path, artifact, skip=valid_chunks)
        from openpilot.common.file_chunker import get_chunk_name
        for i, chunk in enumerate(artifact.chunks):
          chunk_path = get_chunk_name(full_path, i, len(artifact.chunks))
@@ -189,6 +212,17 @@ class ModelManagerSP:
      self._sync_artifact_progress(artifact)
      self._report_status()

+    except DownloadCancelled:
+      # a cancel keeps whatever is on disk: complete chunks resume the next attempt
+      self._download_start_times.pop(artifact.fileName, None)
+      artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
+      artifact.downloadProgress.eta = 0
+      self._sync_artifact_progress(artifact)
+      if self.selected_bundle:
+        self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.failed
+      self._report_status()
+      raise
+
    except Exception as e:
      cloudlog.error(f"Error downloading {filename}: {str(e)}")
      for f in [full_path] + [p for p in (os.path.join(destination_path, f) for f in os.listdir(destination_path)) if filename in p]:
@@ -221,7 +255,6 @@ class ModelManagerSP:
    self.pm.send('modelManagerSP', msg)

  async def _download_bundle(self, model_bundle: custom.ModelManagerSP.ModelBundle, destination_path: str, source: str) -> None:
-    """Downloads a bundle and sets it as the active bundle for its source"""
    self.selected_bundle = model_bundle
    self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.downloading
    for model in self.selected_bundle.models:
@@ -243,6 +276,8 @@ class ModelManagerSP:
          seen_artifacts.add(artifact.fileName)
          await self._process_artifact(artifact, destination_path)

+      if self._download_interrupted():
+        raise DownloadCancelled("Download cancelled")
      self.selected_bundle.status = custom.ModelManagerSP.DownloadStatus.downloaded
      self.params.put(ACTIVE_BUNDLE_KEYS[source], model_bundle.to_dict(), block=True)
      self.active_bundle = get_active_bundle(self.params, usbgpu=self.chestnut_present)
@@ -259,6 +294,27 @@ class ModelManagerSP:
    """Main entry point for downloading a model bundle"""
    asyncio.run(self._download_bundle(model_bundle, destination_path, source))

+  def _process_download_requests(self) -> None:
+    # loops so a ref queued during a download starts in the same tick, without
+    # the bar dropping to idle for a tick between the two transfers
+    last_ref = None
+    while (ref_to_download := self.params.get("ModelManager_DownloadRef")) is not None:
+      if ref_to_download == last_ref:  # a repeating ref falls back to the next tick instead of spinning
+        return
+      last_ref = ref_to_download
+      resolved = resolve_bundle_by_ref(ref_to_download, self.source_models)
+      if not resolved:
+        return
+      model_to_download, source = resolved
+      self._download_ref = ref_to_download
+      try:
+        self.download(model_to_download, Paths.model_root(), source)
+      except Exception as e:
+        cloudlog.exception(e)
+      finally:
+        self._release_download_ref()
+        self.selected_bundle = None
+
  def main_thread(self) -> None:
    """Main thread for model management"""
    rk = Ratekeeper(1, print_delay_threshold=None)
@@ -272,16 +328,7 @@ class ModelManagerSP:
        validate_active_bundles(self.params, self.source_models)
        self.active_bundle = get_active_bundle(self.params, usbgpu=self.chestnut_present)

-        if (ref_to_download := self.params.get("ModelManager_DownloadRef")) is not None:
-          if resolved := resolve_bundle_by_ref(ref_to_download, self.source_models):
-            model_to_download, source = resolved
-            try:
-              self.download(model_to_download, Paths.model_root(), source)
-            except Exception as e:
-              cloudlog.exception(e)
-            finally:
-              self.params.remove("ModelManager_DownloadRef")
-              self.selected_bundle = None
+        self._process_download_requests()

        if self.params.get("ModelManager_ClearCache"):
          self.clear_model_cache()
@@ -305,7 +352,7 @@ class ModelManagerSP:
    for source in ACTIVE_BUNDLE_KEYS:
      if selected_bundle := get_selected_bundle(self.params, source):
        for model in selected_bundle.models:
-          if hasattr(model, 'artifact') and model.artifact.fileName:
+          if model.artifact.fileName:
            active_files.append(model.artifact.fileName)

    # Remove all files except active ones (including their chunk files)
@@ -25,7 +25,9 @@ from openpilot.common.file_chunker import get_chunk_name, get_manifest_path
 from openpilot.selfdrive.test.helpers import http_server_context
 from openpilot.sunnypilot.models import manager as manager_module
 from openpilot.sunnypilot.models.fetcher import ModelFetcher, get_cached_bundles
-from openpilot.sunnypilot.models.helpers import get_active_bundle, get_active_source, get_selected_bundle, resolve_bundle_by_ref
+from openpilot.sunnypilot.models import helpers
+from openpilot.sunnypilot.models.helpers import (get_active_bundle, get_active_source, get_selected_bundle,
+                                                  resolve_bundle_by_ref, validate_active_bundles)
 from openpilot.sunnypilot.models.manager import ModelManagerSP

 CHUNK_BODIES = [b'A' * 5000, b'B' * 5000, b'C' * 3000]
@@ -101,6 +103,7 @@ class ManagerDownloadTestBase(OpenpilotTestCase):
    self.manager = ModelManagerSP.__new__(ModelManagerSP)
    self.manager.params = mock.MagicMock()
    self.manager.params.get.return_value = b'0'  # not cancelled
+    self.manager._download_ref = b'0'
    self.manager.pm = mock.MagicMock()
    self.manager.pm.send.side_effect = self._record_progress
    self.manager.selected_bundle = None
@@ -259,6 +262,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
      artifact = self.make_artifact(chunked=True)
      base_path = os.path.join(self.dest, artifact.fileName)
      self.manager.params.get.side_effect = lambda key: b"ref" if key == "ModelManager_DownloadRef" else None
+      self.manager._download_ref = b"ref"
      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
      assert os.path.isfile(get_manifest_path(base_path))
    self.run_with_server(body)
@@ -277,12 +281,92 @@ class TestManagerDownload(ManagerDownloadTestBase):
        return b"0"

      self.manager.params.get.side_effect = get
+      self.manager._download_ref = b"ref"
      with self.assertRaises(Exception) as ctx:
        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
      assert 'cancelled' in str(ctx.exception).lower()
      assert not os.path.isfile(get_manifest_path(base_path))
    self.run_with_server(body)

+  def test_replaced_download_ref_queues_instead_of_cancelling(self):
+    """Selecting another model mid-transfer lets the running download finish."""
+    def body():
+      artifact = self.make_artifact(chunked=True)
+      base_path = os.path.join(self.dest, artifact.fileName)
+      self.manager.params.get.side_effect = lambda key: b"other-ref" if key == "ModelManager_DownloadRef" else None
+      self.manager._download_ref = b"ref"
+      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
+      assert os.path.isfile(get_manifest_path(base_path))
+    self.run_with_server(body)
+
+  def test_replaced_download_ref_is_kept(self):
+    """A selection made during a download must survive that download's cleanup."""
+    self.manager.params.get.return_value = b"new-ref"
+    self.manager._download_ref = b"old-ref"
+    self.manager._release_download_ref()
+    self.manager.params.remove.assert_not_called()
+
+  def test_own_download_ref_is_released(self):
+    self.manager.params.get.return_value = b"ref"
+    self.manager._download_ref = b"ref"
+    self.manager._release_download_ref()
+    self.manager.params.remove.assert_called_once_with("ModelManager_DownloadRef")
+
+  def test_cached_bundle_cancel_skips_slot_write(self):
+    """A cancel must stop an already-on-disk bundle before it is applied to the slot."""
+    def body():
+      artifact = self.make_artifact(chunked=True)
+      base_path = os.path.join(self.dest, artifact.fileName)
+      for i, data in enumerate(CHUNK_BODIES):
+        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'wb') as f:
+          f.write(data)
+      self._bundle.ref = "test-ref"
+      params, store = self._make_params_with_store()
+      store["ModelManager_DownloadRef"] = None  # removed -> cancelled
+      self.manager.params = params
+      self.manager._download_ref = b"ref"
+      with self.assertRaises(Exception) as ctx:
+        asyncio.run(self.manager._download_bundle(self._bundle, self.dest, "qcom"))
+      assert 'cancelled' in str(ctx.exception).lower()
+      assert "ModelManager_ActiveBundle" not in store
+      assert all(os.path.isfile(p) for p in self.chunk_paths(base_path)), "cancel must not delete cached chunks"
+    self.run_with_server(body)
+
+  def test_resume_skips_valid_chunks(self):
+    """A chunk already on disk is kept and not re-downloaded; progress starts above its share."""
+    def body():
+      artifact = self.make_artifact(chunked=True)
+      base_path = os.path.join(self.dest, artifact.fileName)
+      with open(get_chunk_name(base_path, 0, len(CHUNK_BODIES)), 'wb') as f:
+        f.write(CHUNK_BODIES[0])
+
+      asyncio.run(self.manager._process_artifact(artifact, self.dest))
+
+      chunk0_suffix = get_chunk_name('', 0, len(CHUNK_BODIES))
+      assert not any(p.endswith(chunk0_suffix) for p in DownloadHandler.request_paths), "valid chunk was re-downloaded"
+      for i, expected in enumerate(CHUNK_BODIES):
+        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'rb') as f:
+          assert f.read() == expected
+      assert os.path.isfile(get_manifest_path(base_path))
+      assert min(self.reported) >= (1 / len(CHUNK_BODIES)) * 100 - 1, "progress must not restart below the resumed share"
+    self.run_with_server(body)
+
+  def test_verify_reports_valid_fraction_then_cached(self):
+    """A fully cached bundle publishes climbing verify progress and ends cached."""
+    def body():
+      artifact = self.make_artifact(chunked=True)
+      base_path = os.path.join(self.dest, artifact.fileName)
+      for i, data in enumerate(CHUNK_BODIES):
+        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'wb') as f:
+          f.write(data)
+
+      asyncio.run(self.manager._process_artifact(artifact, self.dest))
+
+      assert DownloadHandler.request_paths == [], "cached bundle must not hit the network"
+      assert [round(p) for p in self.reported[:3]] == [33, 67, 100]
+      assert artifact.downloadProgress.status == custom.ModelManagerSP.DownloadStatus.cached
+    self.run_with_server(body)
+
  def _make_params_with_store(self):
    params = mock.MagicMock()
    store = {}
@@ -302,7 +386,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
    def body():
      artifact = self.make_artifact(chunked=True)
      self._bundle.ref = "test-ref"
-      self._bundle.minimumSelectorVersion = 17
+      self._bundle.minimumSelectorVersion = 18
      params, store = self._make_params_with_store()
      self.manager.params = params
      asyncio.run(self.manager._download_bundle(self._bundle, self.dest, "qcom"))
@@ -322,7 +406,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
    def body():
      self.make_artifact(chunked=True)
      self._bundle.ref = "big-ref"
-      self._bundle.minimumSelectorVersion = 17
+      self._bundle.minimumSelectorVersion = 18
      params, store = self._make_params_with_store()
      self.manager.params = params
      asyncio.run(self.manager._download_bundle(self._bundle, self.dest, "usbgpu"))
@@ -385,7 +469,7 @@ def manifest_bundle(short_name: str, ref: str, index: int = 0, is_big: bool = Fa
    "environment": "release",
    "runner": "tinygrad",
    "is_big": is_big,
-    "minimum_selector_version": "17",
+    "minimum_selector_version": "18",
    "ref": ref,
    "models": [{
      "type": "supercombo",
@@ -532,6 +616,20 @@ class TestSourceCacheIntegrity(OpenpilotTestCase):
    fetch.assert_called_once_with("qcom")
    assert [bundle.ref for bundle in bundles] == ["ddd"]

+  def test_mismatched_refetch_happens_once(self):
+    """If the fresh manifest still fails the source check, the URL is authoritative:
+    trust it instead of refetching at 1 Hz forever."""
+    params = self._make_params({"bundles": [manifest_bundle("big", "bbb", is_big=True)]},
+                               {"bundles": [manifest_bundle("big2", "ccc", is_big=True)]})
+    fetcher = ModelFetcher(params)
+    fetched = self._fetched(manifest_bundle("big", "bbb", is_big=True))
+    with mock.patch.object(fetcher, "_fetch_and_cache_models", return_value=fetched) as fetch:
+      first = fetcher.get_bundles_for_source("qcom")
+      second = fetcher.get_bundles_for_source("qcom")
+    fetch.assert_called_once_with("qcom")
+    assert [bundle.ref for bundle in first] == ["bbb"]
+    assert [bundle.ref for bundle in second] == ["bbb"]
+
  def test_corrupt_cache_is_refetched(self):
    """A cache that fails to parse (e.g. truncated/foreign JSON) must trigger a
    refetch instead of raising every loop and never recovering."""
@@ -545,15 +643,61 @@ class TestSourceCacheIntegrity(OpenpilotTestCase):
    assert [bundle.ref for bundle in bundles] == ["aaa"]


+class TestActiveBundleValidation(OpenpilotTestCase):
+  """Validation is per-slot: a failed fetch (empty bundle list) must not reset a slot,
+  and resetting one slot must not stomp the runner cache derived from the other."""
+
+  def setUp(self):
+    super().setUp()
+    helpers._LAST_VALIDATED_RAW.clear()
+
+  @staticmethod
+  def _raw_bundle(ref: str, runner: int | None = None) -> dict:
+    bundle = custom.ModelManagerSP.ModelBundle.new_message()
+    bundle.ref = ref
+    bundle.minimumSelectorVersion = 18
+    if runner is not None:
+      bundle.runner = runner
+    return bundle.to_dict()
+
+  def _params(self, qcom=None, usbgpu=None):
+    params = mock.MagicMock()
+
+    def get(key, *args, **kwargs):
+      return {"ModelManager_ActiveBundle": qcom, "ModelManager_ActiveBundleUSBGPU": usbgpu}.get(key)
+
+    params.get.side_effect = get
+    return params
+
+  def test_empty_catalog_does_not_reset_slot(self):
+    params = self._params(qcom=self._raw_bundle("small"))
+    with mock.patch("openpilot.sunnypilot.models.helpers.usbgpu_present", return_value=False):
+      validate_active_bundles(params, {"qcom": [], "usbgpu": []})
+    params.remove.assert_not_called()
+
+  def test_reset_recomputes_runner_from_surviving_slot(self):
+    tinygrad = int(custom.ModelManagerSP.Runner.tinygrad)
+    big_raw = self._raw_bundle("big", runner=tinygrad)
+    params = self._params(qcom=self._raw_bundle("gone"), usbgpu=big_raw)
+    catalog = {"qcom": [custom.ModelManagerSP.ModelBundle(**self._raw_bundle("other"))],
+               "usbgpu": [custom.ModelManagerSP.ModelBundle(**big_raw)]}
+    with mock.patch("openpilot.sunnypilot.models.helpers.usbgpu_present", return_value=True):
+      validate_active_bundles(params, catalog)
+    params.remove.assert_called_once_with("ModelManager_ActiveBundle")
+    runner_puts = [call for call in params.put.call_args_list if call.args[0] == "ModelRunnerTypeCache"]
+    assert [call.args[1] for call in runner_puts] == [tinygrad]
+
+
 class TestActiveBundleSelection(OpenpilotTestCase):
-  """The effective active bundle follows the hardware: the usbgpu slot wins when a GPU
-  is present and compiled, otherwise the qcom slot. Each slot keeps its own selection."""
+  """The effective active bundle is the active source's slot: usbgpu when a GPU is
+  present, qcom otherwise. An empty active slot means the hardware default (stock
+  runner), never the other slot's pick - modeld_v2 requires a real bundle."""

  @staticmethod
  def _raw_bundle(ref: str) -> dict:
    bundle = custom.ModelManagerSP.ModelBundle.new_message()
    bundle.ref = ref
-    bundle.minimumSelectorVersion = 17
+    bundle.minimumSelectorVersion = 18
    return bundle.to_dict()

  def _params(self, qcom=None, usbgpu=None):
@@ -584,10 +728,10 @@ class TestActiveBundleSelection(OpenpilotTestCase):
    with mock.patch("openpilot.sunnypilot.models.helpers.usbgpu_present", return_value=True):
      assert get_active_bundle(params).ref == "big"

-  def test_gpu_without_big_selection_falls_back_to_small(self):
+  def test_gpu_without_big_selection_is_hardware_default(self):
    params = self._params(qcom=self._raw_bundle("small"), usbgpu=None)
    with mock.patch("openpilot.sunnypilot.models.helpers.usbgpu_present", return_value=True):
-      assert get_active_bundle(params).ref == "small"
+      assert get_active_bundle(params) is None


 class TestEffectiveSource(OpenpilotTestCase):
@@ -600,7 +744,7 @@ class TestEffectiveSource(OpenpilotTestCase):
  def _raw_bundle(ref: str) -> dict:
    bundle = custom.ModelManagerSP.ModelBundle.new_message()
    bundle.ref = ref
-    bundle.minimumSelectorVersion = 17
+    bundle.minimumSelectorVersion = 18
    return bundle.to_dict()

  def test_runtime_no_gpu(self):
@@ -1,13 +1,11 @@
 import requests

-from openpilot.common.params import Params
 from openpilot.sunnypilot.models.tinygrad_ref import get_tinygrad_ref
 from openpilot.sunnypilot.models.fetcher import ModelFetcher
 from openpilot.common.test import OpenpilotTestCase

 def fetch_tinygrad_ref():
-  fetcher = ModelFetcher(Params())
-  response = requests.get(fetcher.model_url, timeout=10)
+  response = requests.get(ModelFetcher.MODEL_URL, timeout=10)
  response.raise_for_status()
  json_data = response.json()
  return json_data.get("tinygrad_ref")
@@ -84,6 +84,21 @@ def _migrate_tesla_mads_screen_button(_params):
    cloudlog.exception(f"Error migrating TeslaMadsScreenButton: {e}")


+def _migrate_model_bundle_slots(_params):
+  # Pre-split, a chestnut user's big-model selection lived in the single
+  # ActiveBundle. Seed both slots; validation drops whichever does not match
+  # its own manifest.
+  try:
+    if _params.get("ModelManager_ActiveBundleUSBGPU") is not None:
+      return
+    if (bundle := _params.get("ModelManager_ActiveBundle")) is None:
+      return
+    _params.put("ModelManager_ActiveBundleUSBGPU", bundle, block=True)
+    cloudlog.info("params_migration: seeded ModelManager_ActiveBundleUSBGPU from ModelManager_ActiveBundle")
+  except Exception as e:
+    cloudlog.exception(f"Error migrating model bundle slots: {e}")
+
+
 def run_migration(_params):
  # migrate OnroadScreenOffBrightness
  if _params.get("OnroadScreenOffBrightnessMigrated") != ONROAD_BRIGHTNESS_MIGRATION_VERSION:
@@ -120,3 +135,6 @@ def run_migration(_params):

  # seed TeslaMadsScreenButton for existing Tesla installs
  _migrate_tesla_mads_screen_button(_params)
+
+  # seed the usbgpu model slot from the pre-split single slot
+  _migrate_model_bundle_slots(_params)
@@ -0,0 +1,36 @@
+"""
+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.
+"""
+
+from openpilot.common.params import Params
+from openpilot.common.test import OpenpilotTestCase
+from openpilot.sunnypilot.system.params_migration import _migrate_model_bundle_slots
+
+
+class TestModelBundleSlotMigration(OpenpilotTestCase):
+  """Pre-split, a chestnut user's big-model selection lived in the single ActiveBundle.
+  The migration seeds both slots; per-source validation later drops whichever does not
+  match its own manifest."""
+
+  def test_seeds_usbgpu_slot_from_active_bundle(self):
+    params = Params()
+    bundle = {"ref": "big", "minimumSelectorVersion": 18}
+    params.put("ModelManager_ActiveBundle", bundle, block=True)
+    _migrate_model_bundle_slots(params)
+    assert params.get("ModelManager_ActiveBundleUSBGPU") == bundle
+    assert params.get("ModelManager_ActiveBundle") == bundle
+
+  def test_noop_when_usbgpu_slot_already_set(self):
+    params = Params()
+    params.put("ModelManager_ActiveBundle", {"ref": "small"}, block=True)
+    params.put("ModelManager_ActiveBundleUSBGPU", {"ref": "big"}, block=True)
+    _migrate_model_bundle_slots(params)
+    assert params.get("ModelManager_ActiveBundleUSBGPU") == {"ref": "big"}
+
+  def test_noop_when_no_selection(self):
+    params = Params()
+    _migrate_model_bundle_slots(params)
+    assert params.get("ModelManager_ActiveBundleUSBGPU") is None
@@ -8,12 +8,26 @@ from collections.abc import Callable

 import pyray as rl

-from openpilot.system.ui.lib.application import FontWeight
+from openpilot.system.ui.lib.application import gui_app, FontWeight
 from openpilot.system.ui.sunnypilot.lib.styles import style
 from openpilot.system.ui.sunnypilot.widgets.list_view import ButtonActionSP
-from openpilot.system.ui.widgets.label import UnifiedLabel
+from openpilot.system.ui.widgets.label import ScrollState, UnifiedLabel
 from openpilot.system.ui.widgets.list_view import BUTTON_WIDTH, BUTTON_HEIGHT, TEXT_PADDING, _resolve_value

+SCROLL_SPEED = 1.2  # stock is 0.8, boosted 50% to compensate for larger font (50 vs 32)
+SCROLL_REFERENCE_FPS = 60.
+
+
+class UnifiedLabelSP(UnifiedLabel):
+  # stock scroll formula (0.8 / 60 * fps) is inverted — pre-correct so speed is constant px/sec
+  def _render(self, _):
+    if self._needs_scroll and self._scroll_state == ScrollState.SCROLLING:
+      fps = gui_app.target_fps
+      wrong_step = 0.8 / SCROLL_REFERENCE_FPS * fps
+      correct_step = SCROLL_SPEED * SCROLL_REFERENCE_FPS / fps
+      self._scroll_offset -= (correct_step - wrong_step)
+    super()._render(_)
+

 class NoElideButtonAction(ButtonActionSP):
  def get_width_hint(self):
@@ -21,14 +35,12 @@ class NoElideButtonAction(ButtonActionSP):


 class ScrollingButtonAction(ButtonActionSP):
-  """ButtonActionSP whose value scrolls instead of eliding when it doesn't fit."""
-
  def __init__(self, text: str | Callable[[], str], width: int = style.BUTTON_ACTION_WIDTH,
               enabled: bool | Callable[[], bool] = True):
    super().__init__(text=text, width=width, enabled=enabled)
-    self._value_label = UnifiedLabel("", font_size=style.ITEM_TEXT_FONT_SIZE, font_weight=FontWeight.NORMAL,
-                                     text_color=self._value_color, scroll=True,
-                                     alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE)
+    self._value_label = UnifiedLabelSP("", font_size=style.ITEM_TEXT_FONT_SIZE, font_weight=FontWeight.NORMAL,
+                                       text_color=self._value_color, scroll=True,
+                                       alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE)

  def set_value(self, value: str | Callable[[], str], color: rl.Color = style.ITEM_TEXT_VALUE_COLOR):
    if self.value != _resolve_value(value, ""):
@@ -16,6 +16,7 @@ from openpilot.system.ui.lib.text_measure import measure_text_cached
 from openpilot.system.ui.sunnypilot.lib.styles import style
 from openpilot.system.ui.sunnypilot.widgets.list_view import ListItemSP
 from openpilot.system.ui.widgets.label import UnifiedLabel
+from openpilot.system.ui.sunnypilot.lib.utils import UnifiedLabelSP
 from openpilot.system.ui.widgets.list_view import ItemAction

 FONT_SIZE = style.ITEM_TEXT_FONT_SIZE
@@ -24,6 +25,8 @@ ICON_PADDING = 12

 BAR_WIDTH = 1100
 BAR_HEIGHT = 20
+SEGMENT_GAP = 24
+SEGMENT_NAME_MAX_WIDTH = 380
 BAR_GAP = 16
 BAR_RADIUS = BAR_HEIGHT / 2
 CAPSULE_POINTS = 24
@@ -45,6 +48,8 @@ class DownloadStatusAction(ItemAction):
    super().__init__(width=BAR_WIDTH)
    self.name = ""
    self.status_text = ""
+    self.segments: list[tuple[str, rl.Color, str | None, rl.Color | None]] | None = None
+    self._segment_labels: list[UnifiedLabelSP] = []
    self.downloading = False
    self.text_color = rl.GRAY
    self.icon: str | None = None
@@ -62,7 +67,8 @@ class DownloadStatusAction(ItemAction):
                                       alignment=rl.GuiTextAlignment.TEXT_ALIGN_RIGHT,
                                       alignment_vertical=rl.GuiTextAlignmentVertical.TEXT_ALIGN_MIDDLE)

-  def update(self, name, downloading=False, progress=0.0, status_text="", text_color=rl.GRAY, icon=None, icon_color=None):
+  def update(self, name, downloading=False, progress=0.0, status_text="", text_color=rl.GRAY, icon=None, icon_color=None, segments=None):
+    self.segments = segments
    if downloading and not self.downloading:
      self._name_label.reset_shimmer()
      self._progress.x = progress
@@ -85,11 +91,22 @@ class DownloadStatusAction(ItemAction):
  def get_width_hint(self) -> float:
    if self.downloading:
      return BAR_WIDTH
+    if self.segments:
+      return sum(total for _, _, total in self._measured_segments())
    width = measure_text_cached(self._font, self._idle_text, FONT_SIZE).x
    if self.icon:
      width += ICON_SIZE + ICON_PADDING
    return width

+  def _measured_segments(self):
+    """[(segment, text width, total width incl. icon and gap)]"""
+    out = []
+    for i, seg in enumerate(self.segments or []):
+      text_width = min(measure_text_cached(self._font, seg[0], FONT_SIZE).x, SEGMENT_NAME_MAX_WIDTH)
+      total = text_width + (ICON_PADDING + ICON_SIZE if seg[2] else 0) + (SEGMENT_GAP if i else 0)
+      out.append((seg, text_width, total))
+    return out
+
  def _render(self, rect: rl.Rectangle):
    if self.downloading:
      self._render_downloading(rect)
@@ -134,6 +151,8 @@ class DownloadStatusAction(ItemAction):

  def _render_downloading(self, rect: rl.Rectangle):
    percent = f"{int(self._progress.x)}%"
+    if self.status_text:
+      percent = f"{self.status_text} {percent}"
    text_height = measure_text_cached(self._font, percent, FONT_SIZE).y
    top = rect.y + (rect.height - (text_height + BAR_GAP + BAR_HEIGHT)) / 2

@@ -148,6 +167,9 @@ class DownloadStatusAction(ItemAction):
    self._draw_fill(rail, max(0.0, min(rect.width, rect.width * (self._progress.x / 100.0))))

  def _render_idle(self, rect: rl.Rectangle):
+    if self.segments:
+      self._render_segments(rect)
+      return
    text = self._idle_text
    text_size = measure_text_cached(self._font, text, FONT_SIZE)
    right = rect.x + rect.width
@@ -161,6 +183,29 @@ class DownloadStatusAction(ItemAction):
    rl.draw_text_ex(self._font, text, rl.Vector2(right - text_size.x, rect.y + (rect.height - text_size.y) / 2),
                    FONT_SIZE, 0, self.text_color)

+  def _render_segments(self, rect: rl.Rectangle):
+    measured = self._measured_segments()
+    while len(self._segment_labels) < len(measured):
+      self._segment_labels.append(UnifiedLabelSP("", font_size=FONT_SIZE, max_width=SEGMENT_NAME_MAX_WIDTH,
+                                                 scroll=True, wrap_text=False))
+    x = rect.x + rect.width - sum(total for _, _, total in measured)
+    for i, ((text, color, icon, icon_color), text_width, _) in enumerate(measured):
+      if i:
+        x += SEGMENT_GAP
+      label = self._segment_labels[i]
+      if label.text != text:
+        label.set_text(text)
+      label.set_text_color(color)
+      text_height = measure_text_cached(self._font, text, FONT_SIZE).y
+      label.set_position(x, rect.y + (rect.height - text_height) / 2)
+      label.render()
+      x += text_width
+      if icon:
+        texture = gui_app.texture(icon, ICON_SIZE, ICON_SIZE, keep_aspect_ratio=True)
+        rl.draw_texture_v(texture, rl.Vector2(x + ICON_PADDING, rect.y + (rect.height - texture.height) / 2),
+                          icon_color or color)
+        x += ICON_PADDING + ICON_SIZE
+

 def download_status_item(title):
  return ListItemSP(title=title, action_item=DownloadStatusAction(), title_color=style.ITEM_TEXT_COLOR)
@@ -136,7 +136,7 @@ def generate_chunked_model(driving_pkl: Path) -> dict:


 def create_metadata_json(models: list, output_dir: Path, custom_name=None, short_name=None, is_20hz=False, upstream_branch="unknown",
-                         onnx_sha256=None) -> None:
+                         onnx_sha256=None, is_big=False) -> None:
  bundle_json = {
    "short_name": short_name,
    "display_name": custom_name or upstream_branch,
@@ -149,6 +149,7 @@ def create_metadata_json(models: list, output_dir: Path, custom_name=None, short
    "generation": "-1",
    "build_time": datetime.now(UTC).strftime("%Y-%m-%dT%H:%M:%SZ"),
    "overrides": {},
+    "is_big": is_big,
    "models": models,
  }

@@ -186,6 +187,8 @@ if __name__ == "__main__":
    print(f"No driving_tinygrad.pkl found in {_output_dir}", file=sys.stderr)
    sys.exit(1)

+  is_big = _driving_pkl.name.startswith('big_')
+
  if _pkl:
    new_pkl = _output_dir / f"driving_{_pkl}_tinygrad.pkl"
    if not new_pkl.exists():
@@ -196,4 +199,4 @@ if __name__ == "__main__":
  _model_metadata = generate_chunked_model(_driving_pkl)
  _onnx_sha256 = _hash_onnx_files(Path(args.model_dir))
  create_metadata_json([_model_metadata], _output_dir, args.custom_name, _short_name, args.is_20hz, args.upstream_branch,
-                       onnx_sha256=_onnx_sha256)
+                       onnx_sha256=_onnx_sha256, is_big=is_big)
@@ -47,7 +47,7 @@ git rm -rf $OUTPUT_DIR/.git || true # Doing cleanup, but it might fail if the .g
 git remote remove origin || true # ensure cleanup
 git remote add origin $GIT_ORIGIN
 #git push origin -d $DEV_BRANCH || true # Ensuring we delete the remote branch if it exists as we are wiping it out
-git fetch origin $DEV_BRANCH || (git checkout -b $DEV_BRANCH && git commit --allow-empty -m "sunnypilot v$VERSION release" && git push -u origin $DEV_BRANCH)
+git fetch --depth 1 origin $DEV_BRANCH || (git checkout -b $DEV_BRANCH && git commit --allow-empty -m "sunnypilot v$VERSION release" && git push -u origin $DEV_BRANCH)

 echo "[-] committing version $VERSION T=$SECONDS"
 git add -f .