diff --git a/.github/workflows/build-all-tinygrad-models.yaml b/.github/workflows/build-all-tinygrad-models.yaml
index 0eedb04703..1ba407b3cf 100644
--- a/.github/workflows/build-all-tinygrad-models.yaml
+++ b/.github/workflows/build-all-tinygrad-models.yaml
@@ -7,6 +7,19 @@ on:
         description: 'Minimum selector version required for the models (see helpers.py or readme.md)'
         required: true
         type: string
+      target_hardware:
+        description: 'Hardware target to compile for (qcom or usbgpu)'
+        required: true
+        type: choice
+        default: 'qcom'
+        options:
+          - qcom
+          - usbgpu
+      hf_repo:
+        description: 'Hugging Face dataset repository'
+        required: false
+        type: string
+        default: 'sunnypilot/sunnypilot_models_v1'
 
 jobs:
   setup:
@@ -46,13 +59,14 @@ jobs:
         id: get-json
         run: |
           cd docs/docs
-          latest=$(ls driving_models_v*.json | sed -E 's/.*_v([0-9]+)\.json/\1/' | sort -n | tail -1)
+          PREFIX="driving_models_${{ inputs.target_hardware == 'usbgpu' && 'usbgpu_' || '' }}v"
+          latest=$(ls ${PREFIX}*.json | sed -E "s/${PREFIX}([0-9]+)\.json/\1/" | sort -n | tail -1)
           next=$((latest+1))
-          json_file="driving_models_v${next}.json"
-          cp "driving_models_v${latest}.json" "$json_file"
+          json_file="${PREFIX}${next}.json"
+          cp "${PREFIX}${latest}.json" "$json_file"
           echo "json_file=docs/docs/$json_file" >> $GITHUB_OUTPUT
           echo "json_version=$((next+0))" >> $GITHUB_OUTPUT
-          echo "SRC_JSON_FILE=docs/docs/driving_models_v${latest}.json" >> $GITHUB_ENV
+          echo "SRC_JSON_FILE=docs/docs/${PREFIX}${latest}.json" >> $GITHUB_ENV
 
       - name: Extract tinygrad models
         id: set-matrix
@@ -61,45 +75,23 @@ jobs:
           jq -c '[.bundles[] | select(.runner=="tinygrad") | {ref, display_name: (.display_name | gsub(" \\([^)]*\\)"; "")), is_20hz}]' "$(basename "${SRC_JSON_FILE}")" > matrix.json
           echo "model_matrix=$(cat matrix.json)" >> $GITHUB_OUTPUT
 
-      - name: Set up SSH
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.GITLAB_SSH_PRIVATE_KEY }}
-      - run: |
-          mkdir -p ~/.ssh
-          ssh-keyscan -H gitlab.com >> ~/.ssh/known_hosts
-
-      - name: Clone GitLab docs repo and create new recompiled dir
+      - name: Get next recompiled dir number
         id: create-recompiled-dir
         env:
-          GIT_SSH_COMMAND: 'ssh -o UserKnownHostsFile=~/.ssh/known_hosts'
+          HF_REPO: ${{ github.event.inputs.hf_repo }}
         run: |
-          git clone --depth 1 --filter=tree:0 --sparse git@gitlab.com:sunnypilot/public/${{ vars.MODELS_GITLAB }} gitlab_docs
-          cd gitlab_docs
-          git checkout main
-          git sparse-checkout set --no-cone models/
-          cd models
-          latest_dir=$(ls -d recompiled* 2>/dev/null | sed -E 's/recompiled([0-9]+)/\1/' | sort -n | tail -1)
-          if [[ -z "$latest_dir" ]]; then
-            next_dir=1
-          else
-            next_dir=$((latest_dir+1))
-          fi
-          recompiled_dir="${next_dir}"
-          mkdir -p "recompiled${recompiled_dir}"
-          touch "recompiled${recompiled_dir}/.gitkeep"
-          cd ../..
+          pip install huggingface_hub
+          recompiled_dir=$(python3 -c "
+          from huggingface_hub import HfApi
+          import re, sys
+          api = HfApi()
+          files = api.list_repo_files(repo_id=sys.argv[1], repo_type='dataset')
+          dirs = [re.search(r'models/recompiled([0-9]+)', f) for f in files]
+          nums = [int(m.group(1)) for m in dirs if m]
+          print(max(nums) + 1)
+          " "$HF_REPO")
           echo "recompiled_dir=$recompiled_dir" >> $GITHUB_OUTPUT
 
-      - name: Push empty recompiled dir to GitLab
-        run: |
-          cd gitlab_docs
-          git add models/recompiled${{ steps.create-recompiled-dir.outputs.recompiled_dir }}
-          git config --global user.name "GitHub Action"
-          git config --global user.email "action@github.com"
-          git commit -m "Add recompiled${{ steps.create-recompiled-dir.outputs.recompiled_dir }} for build-all" || echo "No changes to commit"
-          git push origin main
-
       - name: Push new JSON to GitHub docs repo
         run: |
           cd docs
@@ -123,25 +115,30 @@ jobs:
       is_20hz: ${{ matrix.model.is_20hz }}
       recompiled_dir: ${{ needs.setup.outputs.recompiled_dir }}
       json_version: ${{ needs.setup.outputs.json_version }}
+      target_hardware: ${{ github.event.inputs.target_hardware }}
+      hf_repo: ${{ github.event.inputs.hf_repo }}
+      set_min_version: ${{ github.event.inputs.set_min_version }}
+      tinygrad_ref: ${{ needs.setup.outputs.tinygrad_ref }}
     secrets: inherit
 
   retry_failed_models:
     needs: [setup, get_and_build]
     runs-on: ubuntu-latest
-    if: ${{ needs.setup.result != 'failure' && !cancelled() }}
+    if: ${{ !cancelled() && needs.setup.result == 'success' && (needs.get_and_build.result == 'success' || needs.get_and_build.result == 'failure') }}
     outputs:
       retry_matrix: ${{ steps.set-retry-matrix.outputs.retry_matrix }}
     steps:
       - uses: actions/download-artifact@v4
         with:
-          pattern: model-*
+          pattern: artifact-name-*
           path: output
+        continue-on-error: true
 
       - id: set-retry-matrix
         run: |
           echo '${{ needs.setup.outputs.model_matrix }}' > matrix.json
-          built=(); while IFS= read -r line; do built+=("$line"); done < <(
-            find output -maxdepth 1 -name 'model-*' -printf "%f\n" | sed -E 's/^model-//' | sed -E 's/-[0-9]+$//' | sed -E 's/ \([^)]*\)//' | awk '{gsub(/^ +| +$/, ""); print}'
+          built=(); while IFS= read -r line; do [ -n "$line" ] && built+=("$line"); done < <(
+            find output -maxdepth 1 -name 'artifact-name-*' -printf "%f\n" 2>/dev/null | sed -E 's/^artifact-name-//' | awk '{gsub(/^ +| +$/, ""); print}'
           )
           jq -c --argjson built "$(printf '%s\n' "${built[@]}" | jq -R . | jq -s .)" \
             'map(select(.display_name as $n | ($built | index($n | gsub("^ +| +$"; "")) | not)))' matrix.json > retry_matrix.json
@@ -149,7 +146,7 @@ jobs:
 
   retry_get_and_build:
     needs: [setup, get_and_build, retry_failed_models]
-    if: ${{ needs.get_and_build.result == 'failure' || (needs.retry_failed_models.outputs.retry_matrix != '[]' && needs.retry_failed_models.outputs.retry_matrix != '') }}
+    if: ${{ !cancelled() && needs.retry_failed_models.result == 'success' && needs.retry_failed_models.outputs.retry_matrix != '[]' && needs.retry_failed_models.outputs.retry_matrix != '' }}
     strategy:
       matrix:
         model: ${{ fromJson(needs.retry_failed_models.outputs.retry_matrix) }}
@@ -161,146 +158,9 @@ jobs:
       is_20hz: ${{ matrix.model.is_20hz }}
       recompiled_dir: ${{ needs.setup.outputs.recompiled_dir }}
       json_version: ${{ needs.setup.outputs.json_version }}
+      target_hardware: ${{ github.event.inputs.target_hardware }}
       artifact_suffix: -retry
+      hf_repo: ${{ github.event.inputs.hf_repo }}
+      set_min_version: ${{ github.event.inputs.set_min_version }}
+      tinygrad_ref: ${{ needs.setup.outputs.tinygrad_ref }}
     secrets: inherit
-
-  publish_models:
-    name: Publish models sequentially
-    needs: [setup, get_and_build, retry_failed_models, retry_get_and_build]
-    if: ${{ !cancelled() && (needs.get_and_build.result != 'failure' || needs.retry_get_and_build.result == 'success' || (needs.retry_failed_models.outputs.retry_matrix != '[]' && needs.retry_failed_models.outputs.retry_matrix != '')) }}
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      max-parallel: 1
-      matrix:
-        model: ${{ fromJson(needs.setup.outputs.model_matrix) }}
-    env:
-      RECOMPILED_DIR: recompiled${{ needs.setup.outputs.recompiled_dir }}
-      JSON_FILE: ${{ needs.setup.outputs.json_file }}
-      ARTIFACT_NAME_INPUT: ${{ matrix.model.display_name }}
-    steps:
-      - name: Set up SSH
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.GITLAB_SSH_PRIVATE_KEY }}
-
-      - name: Add GitLab.com SSH key to known_hosts
-        run: |
-          mkdir -p ~/.ssh
-          ssh-keyscan -H gitlab.com >> ~/.ssh/known_hosts
-
-      - name: Clone GitLab docs repo
-        env:
-          GIT_SSH_COMMAND: 'ssh -o UserKnownHostsFile=~/.ssh/known_hosts'
-        run: |
-          echo "Cloning GitLab"
-          git clone --depth 1 --filter=tree:0 --sparse git@gitlab.com:sunnypilot/public/${{ vars.MODELS_GITLAB }} gitlab_docs
-          cd gitlab_docs
-          echo "checkout models/${RECOMPILED_DIR}"
-          git sparse-checkout set --no-cone models/${RECOMPILED_DIR}
-          git checkout main
-          cd ..
-
-      - name: Checkout docs repo
-        uses: actions/checkout@v4
-        with:
-          repository: sunnypilot/sunnypilot-models
-          ref: gh-pages
-          path: docs
-          ssh-key: ${{ secrets.CI_SUNNYPILOT_DOCS_PRIVATE_KEY }}
-
-      - name: Validate recompiled dir and JSON version
-        run: |
-          if [ ! -d "gitlab_docs/models/$RECOMPILED_DIR" ]; then
-            echo "Recompiled dir $RECOMPILED_DIR does not exist in GitLab repo"
-            exit 1
-          fi
-          if [ ! -f "$JSON_FILE" ]; then
-            echo "JSON file $JSON_FILE does not exist!"
-            exit 1
-          fi
-
-      - name: Download artifact name file
-        uses: actions/download-artifact@v4
-        with:
-          name: artifact-name-${{ env.ARTIFACT_NAME_INPUT }}
-          path: artifact_name
-
-      - name: Read artifact name
-        id: read-artifact-name
-        run: |
-          ARTIFACT_NAME=$(cat artifact_name/artifact_name.txt)
-          echo "artifact_name=$ARTIFACT_NAME" >> $GITHUB_OUTPUT
-
-      - name: Download model artifact
-        uses: actions/download-artifact@v4
-        with:
-          name: ${{ steps.read-artifact-name.outputs.artifact_name }}
-          path: output
-
-      - name: Remove onnx files bc not needed for recompiled dir since they already exist from single build
-        run: |
-          find output -type f -name '*.onnx' -delete
-          find output -type f -name 'big_*.pkl' -delete
-          find output -type f -name 'dmonitoring_model_tinygrad.pkl' -delete
-
-      - name: Copy model artifacts to gitlab
-        env:
-          ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
-        run: |
-          ARTIFACT_DIR="gitlab_docs/models/${RECOMPILED_DIR}/${ARTIFACT_NAME}"
-          mkdir -p "$ARTIFACT_DIR"
-          for path in output/*; do
-            if [ "$(basename "$path")" = "artifact_name.txt" ]; then
-              continue
-            fi
-            name="$(basename "$path")"
-            if [ -d "$path" ]; then
-              mkdir -p "$ARTIFACT_DIR/$name"
-              cp -r "$path"/* "$ARTIFACT_DIR/$name/"
-              echo "Copied dir $name -> $ARTIFACT_DIR/$name"
-            else
-              cp "$path" "$ARTIFACT_DIR/"
-              echo "Copied file $name -> $ARTIFACT_DIR/"
-            fi
-          done
-
-      - name: Push recompiled dir to GitLab
-        env:
-          GITLAB_SSH_PRIVATE_KEY: ${{ secrets.GITLAB_SSH_PRIVATE_KEY }}
-        run: |
-          cd gitlab_docs
-          git checkout main
-          git pull origin main
-          for d in models/"$RECOMPILED_DIR"/*/; do
-            git sparse-checkout add "$d"
-          done
-          git add models/"$RECOMPILED_DIR"
-          git config --global user.name "GitHub Action"
-          git config --global user.email "action@github.com"
-          git commit -m "Update $RECOMPILED_DIR with model from build-all-tinygrad-models" || echo "No changes to commit"
-          git push origin main
-      - run: |
-          cd docs
-          git pull origin gh-pages
-
-      - name: update json
-        run: |
-          ARGS=""
-          [ -n "${{ inputs.set_min_version }}" ] && ARGS="$ARGS --set-min-version \"${{ inputs.set_min_version }}\""
-          ARGS="$ARGS --sort-by-date"
-          ARGS="$ARGS --tinygrad-ref \"${{ needs.setup.outputs.tinygrad_ref }}\""
-          eval python3 docs/json_parser.py \
-            --json-path "$JSON_FILE" \
-            --recompiled-dir "gitlab_docs/models/$RECOMPILED_DIR" \
-            $ARGS
-
-      - name: Push updated json to GitHub
-        run: |
-          cd docs
-          git config --global user.name "GitHub Action"
-          git config --global user.email "action@github.com"
-          git checkout gh-pages
-          git add docs/"$(basename $JSON_FILE)"
-          git commit -m "Update $(basename $JSON_FILE) after recompiling model" || echo "No changes to commit"
-          git push origin gh-pages
diff --git a/.github/workflows/build-single-tinygrad-model.yaml b/.github/workflows/build-single-tinygrad-model.yaml
index cf2d870802..1ad06a54e4 100644
--- a/.github/workflows/build-single-tinygrad-model.yaml
+++ b/.github/workflows/build-single-tinygrad-model.yaml
@@ -29,11 +29,24 @@ on:
         required: false
         type: boolean
         default: true
-      bypass_push:
-        description: 'Bypass pushing to GitLab for build-all'
+      target_hardware:
+        description: 'Hardware target to compile for (qcom or usbgpu)'
         required: false
-        default: true
-        type: boolean
+        type: string
+        default: 'qcom'
+      hf_repo:
+        description: 'Hugging Face dataset repository (e.g. sunnypilot/sunnypilot_models_v1)'
+        required: false
+        type: string
+        default: 'sunnypilot/sunnypilot_models_v1'
+      set_min_version:
+        description: 'Minimum selector version'
+        required: false
+        type: string
+      tinygrad_ref:
+        description: 'Tinygrad reference'
+        required: false
+        type: string
   workflow_dispatch:
     inputs:
       upstream_branch:
@@ -65,8 +78,8 @@ on:
           - None
           - Master Models
           - Release Models
-          - 2025 World Models
           - 2026 World Models
+          - 2026 Deep RL Models
           - Custom Merge Models
           - Other
       custom_model_folder:
@@ -81,9 +94,22 @@ on:
         description: 'Minimum selector version'
         required: false
         type: string
+      target_hardware:
+        description: 'Hardware target to compile for'
+        required: false
+        type: choice
+        default: 'qcom'
+        options:
+          - qcom
+          - usbgpu
+      hf_repo:
+        description: 'Hugging Face dataset repository'
+        required: false
+        type: string
+        default: 'sunnypilot/sunnypilot_models_v1'
 env:
   RECOMPILED_DIR: recompiled${{ inputs.recompiled_dir }}
-  JSON_FILE: docs/docs/driving_models_v${{ inputs.json_version }}.json
+  JSON_FILE: docs/docs/driving_models_${{ inputs.target_hardware == 'usbgpu' && 'usbgpu_v' || 'v' }}${{ inputs.json_version }}.json
 
 jobs:
   build_model:
@@ -93,38 +119,20 @@ jobs:
       custom_name: ${{ inputs.custom_name || inputs.upstream_branch }}
       is_20hz: ${{ inputs.is_20hz }}
       artifact_suffix: ${{ inputs.artifact_suffix }}
+      target_hardware: ${{ inputs.target_hardware }}
     secrets: inherit
 
   publish_model:
-    if: ${{ !inputs.bypass_push && !cancelled() }}
+    if: ${{ !cancelled() && needs.build_model.result == 'success' }}
     concurrency:
-      group: gitlab-push-${{ inputs.recompiled_dir }}
+      group: hf-push-${{ inputs.recompiled_dir }}
       cancel-in-progress: false
     needs: build_model
     runs-on: ubuntu-latest
+    permissions:
+      id-token: write
+      contents: write
     steps:
-      - name: Set up SSH
-        uses: webfactory/ssh-agent@v0.9.0
-        with:
-          ssh-private-key: ${{ secrets.GITLAB_SSH_PRIVATE_KEY }}
-
-      - name: Add GitLab.com SSH key to known_hosts
-        run: |
-          mkdir -p ~/.ssh
-          ssh-keyscan -H gitlab.com >> ~/.ssh/known_hosts
-
-      - name: Clone GitLab docs repo
-        env:
-          GIT_SSH_COMMAND: 'ssh -o UserKnownHostsFile=~/.ssh/known_hosts'
-        run: |
-          echo "Cloning GitLab"
-          git clone --depth 1 --filter=tree:0 --sparse git@gitlab.com:sunnypilot/public/${{ vars.MODELS_GITLAB }} gitlab_docs
-          cd gitlab_docs
-          echo "checkout models/${RECOMPILED_DIR}"
-          git sparse-checkout set --no-cone models/${RECOMPILED_DIR}
-          git checkout main
-          cd ..
-
       - name: Checkout docs repo
         uses: actions/checkout@v4
         with:
@@ -133,16 +141,28 @@ jobs:
           path: docs
           ssh-key: ${{ secrets.CI_SUNNYPILOT_DOCS_PRIVATE_KEY }}
 
-      - name: Validate recompiled dir and JSON version
+      - name: Install huggingface_hub
+        run: pip install --upgrade "huggingface_hub>=0.22.0"
+
+      - name: Validate hf_repo and JSON version
+        env:
+          HF_OIDC_RESOURCE: datasets/${{ inputs.hf_repo }}
         run: |
-          if [ ! -d "gitlab_docs/models/$RECOMPILED_DIR" ]; then
-            echo "Recompiled dir $RECOMPILED_DIR does not exist in GitLab repo"
-            exit 1
-          fi
           if [ ! -f "$JSON_FILE" ]; then
             echo "JSON file $JSON_FILE does not exist!"
             exit 1
           fi
+          python3 -c "
+          import sys
+          from huggingface_hub import HfApi
+          try:
+              api = HfApi()
+              api.repo_info(repo_id=sys.argv[1], repo_type='dataset')
+              print(f'Success: Repo {sys.argv[1]} exists.')
+          except Exception as e:
+              print('HF validation failed:', e)
+              sys.exit(1)
+          " "${{ inputs.hf_repo }}"
 
       - name: Download artifact name file
         uses: actions/download-artifact@v4
@@ -162,49 +182,26 @@ jobs:
           name: ${{ steps.read-artifact-name.outputs.artifact_name }}
           path: output
 
-      - name: Remove unwanted files
-        run: |
-          find output -type f -name 'dmonitoring_model_tinygrad.pkl' -delete
-          find output -type f -name 'dmonitoring_model.onnx' -delete
-
-      - name: Copy model artifact(s) to GitLab recompiled dir
+      - name: Create models folder
         env:
           ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
         run: |
-          ARTIFACT_DIR="gitlab_docs/models/${RECOMPILED_DIR}/${ARTIFACT_NAME}"
-          mkdir -p "$ARTIFACT_DIR"
-          for path in output/*; do
-            if [ "$(basename "$path")" = "artifact_name.txt" ]; then
-              continue
-            fi
-            name="$(basename "$path")"
-            if [ -d "$path" ]; then
-              mkdir -p "$ARTIFACT_DIR/$name"
-              cp -r "$path"/* "$ARTIFACT_DIR/$name/"
-              echo "Copied dir $name -> $ARTIFACT_DIR/$name"
-            else
-              cp "$path" "$ARTIFACT_DIR/"
-              echo "Copied file $name -> $ARTIFACT_DIR/"
-            fi
-          done
+          mkdir -p "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}"
+          cp -r output/* "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/"
+          rm -f "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/artifact_name.txt"
 
-      - name: Push recompiled dir to GitLab
+      - name: Upload to Hugging Face
         env:
-          GITLAB_SSH_PRIVATE_KEY: ${{ secrets.GITLAB_SSH_PRIVATE_KEY }}
+          HF_OIDC_RESOURCE: datasets/${{ inputs.hf_repo }}
+          ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
         run: |
-          cd gitlab_docs
-          git checkout main
-          git pull origin main
-          for d in models/"$RECOMPILED_DIR"/*/; do
-            git sparse-checkout add "$d"
-          done
-          git add models/"$RECOMPILED_DIR"
-          git config --global user.name "GitHub Action"
-          git config --global user.email "action@github.com"
-          git commit -m "Create/Update $RECOMPILED_DIR with new/updated model from build-single-tinygrad-model" || echo "No changes to commit"
-          git push origin main
+          hf upload ${{ inputs.hf_repo }} \
+            output/ \
+            "models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/" \
+            --repo-type=dataset
 
-      - run: |
+      - name: Pull gh-pages
+        run: |
           cd docs
           git pull origin gh-pages
 
@@ -220,9 +217,11 @@ jobs:
           fi
           [ -n "${{ inputs.generation }}" ] && ARGS="$ARGS --generation \"${{ inputs.generation }}\""
           [ -n "${{ inputs.version }}" ] && ARGS="$ARGS --version \"${{ inputs.version }}\""
+          [ -n "${{ inputs.set_min_version }}" ] && ARGS="$ARGS --set-min-version \"${{ inputs.set_min_version }}\""
+          [ -n "${{ inputs.tinygrad_ref }}" ] && ARGS="$ARGS --tinygrad-ref \"${{ inputs.tinygrad_ref }}\""
           eval python3 docs/json_parser.py \
             --json-path "$JSON_FILE" \
-            --recompiled-dir "gitlab_docs/models/$RECOMPILED_DIR" \
+            --recompiled-dir "local_models/$RECOMPILED_DIR" \
             --sort-by-date \
             $ARGS
 
diff --git a/.github/workflows/sunnypilot-build-model.yaml b/.github/workflows/sunnypilot-build-model.yaml
index 2c435e58a0..17981d68e5 100644
--- a/.github/workflows/sunnypilot-build-model.yaml
+++ b/.github/workflows/sunnypilot-build-model.yaml
@@ -80,6 +80,7 @@ jobs:
         with:
           repository: commaai/openpilot
           ref: ${{ inputs.upstream_branch }}
+          fetch-depth: 1
           submodules: recursive
           path: openpilot
 
@@ -89,18 +90,25 @@ jobs:
         with:
           repository: sunnypilot/sunnypilot
           ref: ${{ inputs.upstream_branch }}
+          fetch-depth: 1
           submodules: recursive
           path: openpilot
       - name: Get commit date
         id: commit-date
         run: |
-          cd ${{ github.workspace }}/openpilot
+          cd ${{ github.workspace }}/openpilot/openpilot
           commit_date=$(git log -1 --format=%cd --date=format:'%B %d, %Y')
           echo "model_date=${commit_date}" >> $GITHUB_OUTPUT
           cat $GITHUB_OUTPUT
       - run: |
-          cd ${{ github.workspace }}/openpilot
-          git lfs pull
+          cd ${{ github.workspace }}/openpilot/openpilot
+          if [ "${{ inputs.target_hardware }}" != "usbgpu" ]; then
+            git lfs pull -X "selfdrive/modeld/models/big_*.onnx" -X "selfdrive/modeld/models/dmonitoring_*.onnx"
+            rm -f selfdrive/modeld/models/big_*.onnx selfdrive/modeld/models/dmonitoring_*.onnx
+          else
+            git lfs pull -I "selfdrive/modeld/models/big_*.onnx"
+            find selfdrive/modeld/models -name "*.onnx" ! -name "big_*.onnx" -delete
+          fi
       - name: 'Upload Artifact'
         uses: actions/upload-artifact@v4
         with:
@@ -116,24 +124,10 @@ jobs:
     steps:
       - uses: actions/checkout@v4
         with:
+          fetch-depth: 1
           submodules: recursive
 
       - run: git lfs pull
-      - name: Cache SCons
-        uses: actions/cache@v4
-        with:
-          path: ${{env.SCONS_CACHE_DIR}}
-          key: scons-${{ runner.os }}-${{ runner.arch }}-${{ github.head_ref || github.ref_name }}-model-${{ github.sha }}
-          # Note: GitHub Actions enforces cache isolation between different build sources (PR builds, workflow dispatches, etc.)
-          #   for security. Only caches from the default branch are shared across all builds. This is by design and cannot be overridden.
-          restore-keys: |
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ github.head_ref || github.ref_name }}-model
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ github.head_ref || github.ref_name }}
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ env.MASTER_NEW_BRANCH }}-model
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ env.MASTER_BRANCH }}-model
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ env.MASTER_NEW_BRANCH }}
-            scons-${{ runner.os }}-${{ runner.arch }}-${{ env.MASTER_BRANCH }}
-            scons-${{ runner.os }}-${{ runner.arch }}
 
       - name: Set environment variables
         id: set-env
@@ -144,7 +138,7 @@ jobs:
           export UV_PYTHON_PREFERENCE=managed
           export UV_PYTHON_INSTALL_DIR=${HOME}/uv/python
           export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
-          uv sync
+          uv sync --frozen
           printenv >> $GITHUB_ENV
           if [[ "${{ runner.debug }}" == "1" ]]; then
             cat $GITHUB_OUTPUT
@@ -173,8 +167,6 @@ jobs:
         with:
           name: models-${{ env.REF }}${{ inputs.artifact_suffix }}
           path: ${{ env.MODELS_DIR }}
-      - run: |
-          rm -f ${{ env.MODELS_DIR }}/{dmonitoring_model,big_driving_policy,big_driving_vision,big_driving_supercombo}.onnx
 
       - name: Build Model
         run: |
@@ -191,7 +183,7 @@ jobs:
           if [ "${{ inputs.target_hardware }}" == "usbgpu" ]; then
             echo "USBGPU build"
             export USBGPU=1
-            TG_FLAGS="DEV=AMD USBGPU=1 IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+            TG_FLAGS="DEBUG=2 DEV=USB+AMD:LLVM WARP_DEV=QCOM FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2"
             OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
           else
             echo "QCOM build"
@@ -254,10 +246,8 @@ jobs:
           # Copy the model files
           rsync -avm \
             --include='*.dlc' \
-            --include='*.pkl' \
             --include='*.chunk*' \
             --include='*.chunkmanifest' \
-            --include='*.onnx' \
             --exclude='*' \
             --delete-excluded \
             --chown=comma:comma \
diff --git a/openpilot/common/params_keys.h b/openpilot/common/params_keys.h
index bb7989381b..964227e785 100644
--- a/openpilot/common/params_keys.h
+++ b/openpilot/common/params_keys.h
@@ -195,11 +195,14 @@ inline static std::unordered_map<std::string, ParamKeyAttributes> keys = {
 
     // Model Manager params
     {"ModelManager_ActiveBundle", {PERSISTENT, JSON}},
+    {"ModelManager_ActiveJson", {CLEAR_ON_MANAGER_START, STRING}},
     {"ModelManager_ClearCache", {CLEAR_ON_MANAGER_START, BOOL}},
     {"ModelManager_DownloadIndex", {CLEAR_ON_MANAGER_START | CLEAR_ON_ONROAD_TRANSITION, INT}},
     {"ModelManager_Favs", {PERSISTENT | BACKUP, STRING}},
     {"ModelManager_LastSyncTime", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, INT, "0"}},
+    {"ModelManager_LastSyncTime_USBGPU", {CLEAR_ON_MANAGER_START | CLEAR_ON_OFFROAD_TRANSITION, INT, "0"}},
     {"ModelManager_ModelsCache", {PERSISTENT | BACKUP, JSON}},
+    {"ModelManager_ModelsCache_USBGPU", {PERSISTENT | BACKUP, JSON}},
 
     // Neural Network Lateral Control
     {"NeuralNetworkLateralControl", {PERSISTENT | BACKUP, BOOL, "0"}},
diff --git a/openpilot/sunnypilot/SConscript b/openpilot/sunnypilot/SConscript
index 09ad39ab43..587deea5ff 100644
--- a/openpilot/sunnypilot/SConscript
+++ b/openpilot/sunnypilot/SConscript
@@ -1,3 +1,2 @@
 SConscript(['common/transformations/SConscript'])
-SConscript(['modeld_v2/SConscript'])
 SConscript(['selfdrive/locationd/SConscript'])
diff --git a/openpilot/sunnypilot/modeld_v2/SConscript b/openpilot/sunnypilot/modeld_v2/SConscript
deleted file mode 100644
index daaa199ea9..0000000000
--- a/openpilot/sunnypilot/modeld_v2/SConscript
+++ /dev/null
@@ -1,84 +0,0 @@
-import os
-import glob
-
-from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
-from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE
-from openpilot.common.hardware import HARDWARE, PC
-
-Import('env', 'arch', 'release')
-lenv = env.Clone()
-tinygrad_files = ["#"+x for x in glob.glob(env.Dir("#tinygrad_repo").relpath + "/**", recursive=True, root_dir=env.Dir("#").abspath) if 'pycache' not in x]
-
-
-def get_camera_configs():
-  DEVICE_RESOLUTIONS = {
-    "tici": (_ar_ox_fisheye.width, _ar_ox_fisheye.height),
-    "tizi": (_ar_ox_fisheye.width, _ar_ox_fisheye.height),
-    "mici": (_os_fisheye.width, _os_fisheye.height),
-  }
-  if release or PC or 'CI' in os.environ:
-    return set(DEVICE_RESOLUTIONS.values())
-  return [DEVICE_RESOLUTIONS[HARDWARE.get_device_type()]]
-
-CAMERA_CONFIGS = get_camera_configs()
-
-tg_flags = {
-  'larch64': 'DEV=QCOM FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0',
-  'Darwin': f'DEV=CPU HOME={os.path.expanduser("~")}',
-}.get(arch, 'DEV=CPU:LLVM')
-
-image_flag = {
-  'larch64': 'IMAGE=2',
-}.get(arch, 'IMAGE=0')
-
-model_w, model_h = MEDMODEL_INPUT_SIZE
-from openpilot.selfdrive.modeld.constants import ModelConstants
-frame_skip = ModelConstants.MODEL_RUN_FREQ // ModelConstants.MODEL_CONTEXT_FREQ
-camera_res_args = ' '.join(f'{cw}x{ch}' for cw, ch in CAMERA_CONFIGS)
-
-pythonpath_string = 'PYTHONPATH="${PYTHONPATH}:' + env.Dir("#tinygrad_repo").abspath + ':' + env.Dir("#").abspath + '"'
-compile_modeld_script = File("compile_modeld.py").abspath
-upstream_compile_script = File(Dir("#openpilot/selfdrive/modeld").File("compile_modeld.py").abspath)
-script_deps = [File("compile_modeld.py"), upstream_compile_script]
-
-def compile_combined(model_type, onnx_args, output_name):
-  output_pkl = File(f"models/{output_name}").abspath
-  cmd = (f'{pythonpath_string} {tg_flags} {image_flag} python3 {compile_modeld_script} '
-         f'--model-type {model_type} '
-         f'--model-size {model_w}x{model_h} '
-         f'--camera-resolutions {camera_res_args} '
-         f'{onnx_args} '
-         f'--frame-skip {frame_skip} '
-         f'--output {output_pkl}')
-  onnx_files = [f for f in onnx_args.split() if f.endswith('.onnx')]
-  return lenv.Command(output_pkl, tinygrad_files + script_deps + [File(f) for f in onnx_files if os.path.isfile(f)], cmd)
-
-# Vision + Policy (stock default model)
-vision_onnx = File("models/driving_vision.onnx").abspath
-policy_onnx = File("models/driving_policy.onnx").abspath
-if os.path.isfile(vision_onnx) and os.path.isfile(policy_onnx):
-  compile_combined('vision_policy',
-                   f'--vision-onnx {vision_onnx} --policy-onnx {policy_onnx}',
-                   'driving_combined_tinygrad.pkl')
-
-# Vision + Off-Policy
-off_policy_onnx = File("models/driving_off_policy.onnx").abspath
-if os.path.isfile(vision_onnx) and os.path.isfile(off_policy_onnx):
-  policy_arg = f'--policy-onnx {policy_onnx}' if os.path.isfile(policy_onnx) else ''
-  compile_combined('vision_multi_policy',
-                   f'--vision-onnx {vision_onnx} {policy_arg} --off-policy-onnx {off_policy_onnx}',
-                   'driving_combined_multi_tinygrad.pkl')
-
-# Vision + On-Policy + Off-Policy
-on_policy_onnx = File("models/driving_on_policy.onnx").abspath
-if os.path.isfile(vision_onnx) and os.path.isfile(on_policy_onnx) and os.path.isfile(off_policy_onnx):
-  compile_combined('vision_multi_policy',
-                   f'--vision-onnx {vision_onnx} --off-policy-onnx {off_policy_onnx} --on-policy-onnx {on_policy_onnx}',
-                   'driving_combined_tri_tinygrad.pkl')
-
-# Supercombo
-supercombo_onnx = File("models/supercombo.onnx").abspath
-if os.path.isfile(supercombo_onnx):
-  compile_combined('supercombo',
-                   f'--supercombo-onnx {supercombo_onnx}',
-                   'driving_combined_supercombo_tinygrad.pkl')
diff --git a/openpilot/sunnypilot/modeld_v2/compile_modeld.py b/openpilot/sunnypilot/modeld_v2/compile_modeld.py
index def54a4599..4397209005 100755
--- a/openpilot/sunnypilot/modeld_v2/compile_modeld.py
+++ b/openpilot/sunnypilot/modeld_v2/compile_modeld.py
@@ -8,10 +8,10 @@ See the LICENSE.md file in the root directory for more details.
 
 import argparse
 import os
-import pickle
+import tempfile
 import time
-from collections import defaultdict
 from functools import partial
+from openpilot.selfdrive.modeld.helpers import dump_oob, load_oob
 import numpy as np
 os.environ['GMMU'] = '0'
 
@@ -38,6 +38,9 @@ from tinygrad.engine.jit import TinyJit
 from tinygrad.tensor import Tensor
 
 MODEL_TYPES = ('vision_policy', 'supercombo', 'vision_multi_policy')
+WARP_INPUTS = ['tfm', 'big_tfm']
+POLICY_INPUTS = ['img_q', 'big_img_q', 'feat_q', 'desire_q', 'packed_npy_inputs']
+WARP_DEV = os.getenv('WARP_DEV')
 
 
 def _detect_desire_key(shapes: dict) -> str | None:
@@ -76,7 +79,7 @@ def get_policy_npy_shapes(input_shapes: dict, is_supercombo: bool = False) -> tu
 
 
 def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT,
-                            is_supercombo: bool = False, use_packed: bool = True) -> tuple[dict, dict]:
+                            is_supercombo: bool = False) -> tuple[dict, dict]:
   road_key, _ = _detect_vision_keys(input_shapes)
   if not road_key:
     raise ValueError("Vision road key missing from input shapes.")
@@ -92,74 +95,75 @@ def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = D
   desire_shape = input_shapes[desire_key]
   features_buffer = input_shapes.get('features_buffer')
 
-  if use_packed:  # remove packed detection block after all models are recompiled
-    npy_arrays = {
-      'tfm': np.zeros((3, 3), dtype=np.float32),
-      'big_tfm': np.zeros((3, 3), dtype=np.float32)
-    }
+  npy_arrays = {
+    'tfm': np.zeros((3, 3), dtype=np.float32),
+    'big_tfm': np.zeros((3, 3), dtype=np.float32)
+  }
 
-    shapes, sizes = get_policy_npy_shapes(input_shapes, is_supercombo=is_supercombo)
-    packed_npy_inputs = np.zeros(sum(sizes), dtype=np.float32)
+  shapes, sizes = get_policy_npy_shapes(input_shapes, is_supercombo=is_supercombo)
+  packed_npy_inputs = np.zeros(sum(sizes), dtype=np.float32)
 
-    split_indices = np.cumsum(sizes[:-1]) if len(sizes) > 1 else []
-    split_views = np.split(packed_npy_inputs, split_indices) if len(sizes) > 0 else []
-    for (k, s), v in zip(shapes.items(), split_views, strict=True):
-      npy_arrays[k] = v.reshape(s)
+  split_indices = np.cumsum(sizes[:-1]) if len(sizes) > 1 else []
+  split_views = np.split(packed_npy_inputs, split_indices) if len(sizes) > 0 else []
+  for (k, s), v in zip(shapes.items(), split_views, strict=True):
+    npy_arrays[k] = v.reshape(s)
 
-    queues = {
-      'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
-      'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
-      'desire_q': Tensor(np.zeros((frame_skip * desire_shape[1], desire_shape[0], desire_shape[2]),
-                    dtype=np.float32), device=device).contiguous().realize(),
-      'packed_npy_inputs': Tensor(packed_npy_inputs, device='NPY').realize(),
-    }
+  queues = {
+    'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
+    'desire_q': Tensor(np.zeros((frame_skip * desire_shape[1], desire_shape[0], desire_shape[2]),
+                  dtype=np.float32), device=device).contiguous().realize(),
+    'packed_npy_inputs': Tensor(packed_npy_inputs, device='NPY').realize(),
+  }
 
-    if features_buffer:
-      queues['feat_q'] = Tensor(np.zeros((frame_skip * (features_buffer[1] - 1) + 1, features_buffer[0], features_buffer[2]),
-                         dtype=np.float32), device=device).contiguous().realize()
+  if features_buffer:
+    queues['feat_q'] = Tensor(np.zeros((frame_skip * (features_buffer[1] - 1) + 1, features_buffer[0], features_buffer[2]),
+                       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')})
-  else:
-    # TODO-SP: Remove legacy queuing fallback else block after all models are recompiled
-    npy_arrays = {
-      'desire': np.zeros(desire_shape[2], dtype=np.float32),
-      'tfm': np.zeros((3, 3), dtype=np.float32),
-      'big_tfm': np.zeros((3, 3), dtype=np.float32)
-    }
-
-    for key, shape in input_shapes.items():
-      if key not in npy_arrays and 'img' not in key and key not in ('features_buffer', desire_key):
-        npy_arrays[key] = np.zeros(shape, dtype=np.float32)
-
-    queues = {
-      'img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
-      'big_img_q': Tensor(np.zeros(img_buf_shape, dtype=np.uint8), device=device).contiguous().realize(),
-      'desire_q': Tensor(np.zeros((frame_skip * desire_shape[1], desire_shape[0], desire_shape[2]),
-                    dtype=np.float32), device=device).contiguous().realize()
-    }
-
-    if features_buffer:
-      queues['feat_q'] = Tensor(np.zeros((frame_skip * (features_buffer[1] - 1) + 1, features_buffer[0], features_buffer[2]),
-                         dtype=np.float32), device=device).contiguous().realize()
-
-    queues.update({key: Tensor(value, device='NPY').realize() for key, value in npy_arrays.items()})
+  queues.update({key: Tensor(value, device='NPY').realize() for key, value in npy_arrays.items() if key in ('tfm', 'big_tfm')})
 
   return queues, npy_arrays
 
 
 def make_split_input_queues(vision_input_shapes: dict, policy_input_shapes: dict,
-                            frame_skip: int, device: str = Device.DEFAULT, use_packed: bool = True) -> tuple[dict, dict]:
-  return generate_queues_and_npy({**vision_input_shapes, **policy_input_shapes}, frame_skip, device, is_supercombo=False, use_packed=use_packed)
+                            frame_skip: int, device: str = Device.DEFAULT) -> tuple[dict, dict]:
+  return generate_queues_and_npy({**vision_input_shapes, **policy_input_shapes}, frame_skip, device, is_supercombo=False)
 
 
 def make_supercombo_input_queues(input_shapes: dict, frame_skip: int,
-                                 device: str = Device.DEFAULT, use_packed: bool = True) -> tuple[dict, dict]:
-  return generate_queues_and_npy(input_shapes, frame_skip, device, is_supercombo=True, use_packed=use_packed)
+                                 device: str = Device.DEFAULT) -> tuple[dict, dict]:
+  return generate_queues_and_npy(input_shapes, frame_skip, device, is_supercombo=True)
 
 
-def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, model_size: tuple[int, int],
-            features_slice: slice, frame_skip: int, input_shapes: dict, prepare_only: bool):
-  frame_prepare = make_frame_prepare(nv12, *model_size)
+def make_random_images(keys, shape, device):
+  return {k: Tensor.randint(shape, low=0, high=256, dtype=dtypes.uint8, device=device).realize() for k in keys}
+
+
+def make_warp_queues(device=Device.DEFAULT):
+  npy = {
+    'tfm': np.zeros((3, 3), dtype=np.float32),
+    'big_tfm': np.zeros((3, 3), dtype=np.float32),
+  }
+  queues = {k: Tensor(v, device='NPY').realize() for k, v in npy.items()}
+  return queues, npy
+
+
+def make_warp(nv12: NV12Frame, model_w: int, model_h: int):
+  frame_prepare = make_frame_prepare(nv12, model_w, model_h)
+  WARP_DEV = os.getenv('WARP_DEV', Device.DEFAULT)
+
+  def warp(tfm, big_tfm, frame, big_frame):
+    tfm = tfm.to(WARP_DEV)
+    big_tfm = big_tfm.to(WARP_DEV)
+    Tensor.realize(tfm, big_tfm)
+
+    warped_frame = frame_prepare(frame, tfm).unsqueeze(0)
+    warped_big_frame = frame_prepare(big_frame, big_tfm).unsqueeze(0)
+    return Tensor.cat(warped_frame, warped_big_frame)
+  return warp
+
+
+def make_run_policy(vision_runner, policy_runners: list, features_slice: slice, frame_skip: int, input_shapes: dict):
   sample_skip_fn = partial(sample_skip, frame_skip=frame_skip)
   sample_desire_fn = partial(sample_desire, frame_skip=frame_skip)
 
@@ -172,20 +176,14 @@ def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, mode
   is_supercombo = vision_runner is None
   npy_shapes, npy_sizes = get_policy_npy_shapes(input_shapes, is_supercombo=is_supercombo)
 
-  def runner(img_q, big_img_q, feat_q, packed_npy_inputs, frame, big_frame, tfm, big_tfm, **kwargs):
+  def run_policy(warped, img_q, big_img_q, feat_q, packed_npy_inputs, **kwargs):
     desire_q = kwargs['desire_q']
-
     packed_npy_inputs_dev = packed_npy_inputs.to(Device.DEFAULT)
-    tfm_dev = tfm.to(Device.DEFAULT)
-    big_tfm_dev = big_tfm.to(Device.DEFAULT)
+    warped_dev = warped.to(Device.DEFAULT)
+    Tensor.realize(packed_npy_inputs_dev, warped_dev)
 
-    Tensor.realize(packed_npy_inputs_dev, tfm_dev, big_tfm_dev)
-
-    img = shift_and_sample(img_q, frame_prepare(frame, tfm_dev).unsqueeze(0), sample_skip_fn).realize()
-    big_img = shift_and_sample(big_img_q, frame_prepare(big_frame, big_tfm_dev).unsqueeze(0), sample_skip_fn).realize()
-
-    if prepare_only:
-      return img, big_img
+    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()
 
     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))
@@ -220,42 +218,52 @@ def create_jit_runner(vision_runner, policy_runners: list, nv12: NV12Frame, mode
       shift_and_sample(feat_q, new_feat, sample_skip_fn).realize()
     return policy_out
 
-  return runner
+  return run_policy
 
 
-def compile_and_warmup(nv12: NV12Frame, model_size: tuple[int, int], prepare_only: bool, frame_skip: int, vision_runner, policy_runners: list, metadata: dict):
-  print(f"Compiling combined JIT for {nv12.width}x{nv12.height} (prepare_only={prepare_only})...")
+def compile_jit(jit, make_random_inputs, input_keys, make_queues):
+  SEED = 42
+  def random_inputs_run(fn, seed, test_val=None, test_buffers=None, expect_match=True):
+    input_queues, npy = make_queues(Device.DEFAULT)
+    rng = np.random.default_rng(seed)
+    Tensor.manual_seed(seed)
 
-  all_shapes = {key: value for meta in metadata.values() for key, value in meta['input_shapes'].items()}
+    testing = test_val is not None or test_buffers is not None
+    n_runs = 1 if testing else 3
 
-  feat_meta = metadata.get('vision') or metadata.get('model') or metadata.get('policy')
-  if not feat_meta:
-    raise ValueError("Could not find vision, model, or policy metadata.")
+    for i in range(n_runs):
+      for v in npy.values():
+        v[:] = rng.standard_normal(v.shape).astype(v.dtype)
+      Device.default.synchronize()
+      random_inputs = make_random_inputs()
+      st = time.perf_counter()
+      outs = fn(**{k: input_queues[k] for k in input_keys if k in input_queues}, **random_inputs)
+      mt = time.perf_counter()
+      Device.default.synchronize()
+      et = time.perf_counter()
+      print(f"  [{i+1}/{n_runs}] enqueue {(mt-st)*1e3:6.2f} ms -- total {(et-st)*1e3:6.2f} ms")
 
-  features_slice = feat_meta['output_slices']['hidden_state']
-  WARP_DEV = 'CPU' if "USBGPU" in os.environ else Device.DEFAULT
+      if i == 0:
+        val = [np.copy(v.numpy()) for v in (outs if isinstance(outs, tuple) else [outs])] if outs is not None else []
+        buffers = [np.copy(v.numpy().copy()) for v in input_queues.values()]
 
-  is_supercombo = vision_runner is None
-  run_func = create_jit_runner(vision_runner, policy_runners, nv12, model_size, features_slice, frame_skip, all_shapes, prepare_only)
-  run_jit = TinyJit(run_func, prune=True)
-  queues, npy_arrays = generate_queues_and_npy(all_shapes, frame_skip, Device.DEFAULT, is_supercombo=is_supercombo)
+    if test_val is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(val, test_val, strict=True))
+      assert match == expect_match, f"outputs {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    if test_buffers is not None:
+      match = all(np.array_equal(a, b) for a, b in zip(buffers, test_buffers, strict=True))
+      assert match == expect_match, f"buffers {'differ from' if expect_match else 'match'} baseline (seed={seed})"
+    return val, buffers
 
-  for i in range(3):
-    rng = np.random.default_rng(42 + i)
-    frame = Tensor.randint(nv12.size, low=0, high=256, dtype=dtypes.uint8, device=WARP_DEV).realize()
-    big_frame = Tensor.randint(nv12.size, low=0, high=256, dtype=dtypes.uint8, device=WARP_DEV).realize()
-    for arr in npy_arrays.values():
-      arr[:] = rng.standard_normal(arr.shape).astype(arr.dtype)
-
-    Device.default.synchronize()
-    start_time = time.perf_counter()
-    run_jit(**queues, frame=frame, big_frame=big_frame)
-    mid_time = time.perf_counter()
-    Device.default.synchronize()
-    print(f"  [{i + 1}/3] enqueue {(mid_time - start_time) * 1e3:6.2f} ms -- total {(time.perf_counter() - start_time) * 1e3:6.2f} ms")
-
-  # TODO-SP: switch to dump_oob/load_oob on next full recompile of all models
-  return pickle.loads(pickle.dumps(run_jit)) if not prepare_only else run_jit
+  print('capture + replay')
+  test_val, test_buffers = random_inputs_run(jit, SEED)
+  print('pickle round trip')
+  with tempfile.TemporaryFile(dir=".") as f:
+    dump_oob(jit, f)
+    f.seek(0)
+    deserialized_jit = load_oob(f)
+  random_inputs_run(deserialized_jit, SEED, test_val=test_val, test_buffers=test_buffers)
+  return deserialized_jit
 
 
 def _parse_size(size_str: str) -> tuple[int, int]:
@@ -277,19 +285,6 @@ def read_file_chunked_to_shm(path):
   return shm_path
 
 
-def _compile_for_resolutions(camera_resolutions: list, model_size: tuple[int, int], frame_skip: int,
-               vision_runner, policy_runners: list, metadata: dict) -> dict:
-  from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-  return {
-    (cam_w, cam_h): {
-      name: compile_and_warmup(NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h)), model_size, prepare_only,
-                   frame_skip, vision_runner, policy_runners, metadata)
-      for name, prepare_only in [('warp_enqueue', True), ('run_policy', False)]
-    }
-    for cam_w, cam_h in camera_resolutions
-  }
-
-
 def _load_policy_runners(args: argparse.Namespace) -> tuple[list, list]:
   runners, keys = [], []
   for name, onnx_arg in [('policy', args.policy_onnx), ('off_policy', args.off_policy_onnx), ('on_policy', args.on_policy_onnx)]:
@@ -300,7 +295,18 @@ def _load_policy_runners(args: argparse.Namespace) -> tuple[list, list]:
 
 
 if __name__ == "__main__":
+  if 'USB' in os.getenv('DEV', '') or os.getenv('USBGPU'):
+    from openpilot.system.hardware.chestnut.flash import link_up
+    for _ in range(10):
+      if link_up():
+        break
+      time.sleep(1)
+    else:
+      raise RuntimeError("Chestnut not ready, skipping big model build")
+
+  from openpilot.common.file_chunker import chunk_file, get_chunk_targets
   from openpilot.selfdrive.modeld.get_model_metadata import make_metadata_dict
+  from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
   from tinygrad.nn.onnx import OnnxRunner
 
   parser = argparse.ArgumentParser(description="Compile combined JIT pkl for sunnypilot modeld_v2")
@@ -317,7 +323,8 @@ if __name__ == "__main__":
   parser.add_argument('--supercombo-onnx', help='supercombo ONNX (for supercombo)')
 
   args = parser.parse_args()
-  output_data = defaultdict(dict)
+  model_w, model_h = args.model_size
+  output_data = {}
 
   args.vision_onnx = read_file_chunked_to_shm(args.vision_onnx)
   args.policy_onnx = read_file_chunked_to_shm(args.policy_onnx)
@@ -348,17 +355,31 @@ if __name__ == "__main__":
   vision_meta = output_data['metadata'].get('vision', {})
 
   derived_frame_skip = args.frame_skip or derive_frame_skip(vision_meta.get('input_shapes', {}), first_policy_meta.get('input_shapes', {}))
-  output_data.update(_compile_for_resolutions(args.camera_resolutions, args.model_size, derived_frame_skip,
-                        vision_runner, policy_runners, output_data['metadata']))
+  all_shapes = {key: value for meta in output_data['metadata'].values() for key, value in meta['input_shapes'].items()}
+  feat_meta = output_data['metadata'].get('vision') or output_data['metadata'].get('model') or output_data['metadata'].get('policy')
+  assert feat_meta is not None
+  features_slice = feat_meta['output_slices']['hidden_state']
+  is_supercombo = vision_runner is None
+
+  print(f"Compiling run_policy JIT (model_size={model_w}x{model_h}, frame_skip={derived_frame_skip})...")
+  run_policy_func = make_run_policy(vision_runner, policy_runners, features_slice, derived_frame_skip, all_shapes)
+  run_policy_jit = TinyJit(run_policy_func, prune=True)
+  make_policy_queues = partial(generate_queues_and_npy, all_shapes, derived_frame_skip, is_supercombo=is_supercombo)
+  make_random_model_inputs = partial(make_random_images, keys=['warped'], shape=(2, 6, model_h // 2, model_w // 2), device=WARP_DEV)
+  output_data['run_policy'] = compile_jit(run_policy_jit, make_random_model_inputs, POLICY_INPUTS, make_policy_queues)
+
+  for cam_w, cam_h in args.camera_resolutions:
+    print(f"Compiling warp JIT for {cam_w}x{cam_h}...")
+    nv12 = NV12Frame(cam_w, cam_h, *get_nv12_info(cam_w, cam_h))
+    make_random_warp_inputs = partial(make_random_images, keys=['frame', 'big_frame'], shape=nv12.size, device=WARP_DEV)
+    warp = TinyJit(make_warp(nv12, model_w, model_h), prune=True)
+    output_data[(cam_w, cam_h)] = compile_jit(warp, make_random_warp_inputs, WARP_INPUTS, make_warp_queues)
 
   with open(args.output, "wb") as file:
-    # TODO-SP: switch to dump_oob from openpilot/selfdrive/helpers on next full recompile of all models
-    pickle.dump(output_data, file)
+    dump_oob(output_data, file)
 
   pkl_size = os.path.getsize(args.output)
   print(f"Saved combined JIT to {args.output} ({pkl_size / 1e6:.2f} MB)")
-
-  from openpilot.common.file_chunker import chunk_file, get_chunk_targets
   chunk_targets = get_chunk_targets(args.output, pkl_size)
   chunk_file(args.output, chunk_targets)
   print(f"Chunked into {len(chunk_targets) - 1} file(s)")
diff --git a/openpilot/sunnypilot/modeld_v2/modeld.py b/openpilot/sunnypilot/modeld_v2/modeld.py
index 66b560802b..9f3d709537 100755
--- a/openpilot/sunnypilot/modeld_v2/modeld.py
+++ b/openpilot/sunnypilot/modeld_v2/modeld.py
@@ -9,17 +9,13 @@ See the LICENSE.md file in the root directory for more details.
 import os
 os.environ['GMMU'] = '0'
 from openpilot.common.hardware import COMMA_HARDWARE
-os.environ['DEV'] = 'QCOM' if COMMA_HARDWARE else 'CPU'
-USBGPU = "USBGPU" in os.environ
-if USBGPU:
-  os.environ['DEV'] = 'AMD'
-  os.environ['AMD_IFACE'] = 'USB'
-import pickle
+from openpilot.selfdrive.modeld.helpers import usbgpu_present, load_oob
 import time
 import numpy as np
 import openpilot.cereal.messaging as messaging
 from openpilot.cereal import log
 from opendbc.car.structs import car
+from openpilot.cereal.services import SERVICE_LIST
 from setproctitle import setproctitle
 from openpilot.cereal.messaging import PubMaster, SubMaster
 from openpilot.cereal.visionipc import VisionStreamType
@@ -27,7 +23,6 @@ from msgq.visionipc import VisionIpcClient, VisionBuf
 from opendbc.car.car_helpers import get_demo_car_params
 
 from tinygrad.tensor import Tensor
-from tinygrad.device import Device
 
 from openpilot.common.file_chunker import open_file_chunked
 from openpilot.common.swaglog import cloudlog
@@ -40,12 +35,13 @@ from openpilot.system import sentry
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, smooth_value
+from openpilot.selfdrive.modeld.modeld import ChestnutState
 
 from openpilot.sunnypilot.modeld_v2.fill_model_msg import fill_model_msg, fill_pose_msg, PublishState, get_curvature_from_output
 from openpilot.sunnypilot.modeld_v2.constants import Plan
 from openpilot.sunnypilot.modeld_v2.meta_helper import load_meta_constants
 from openpilot.sunnypilot.modeld_v2.camera_offset_helper import CameraOffsetHelper
-from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues
+from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, make_split_input_queues, make_supercombo_input_queues, WARP_INPUTS, POLICY_INPUTS
 
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
@@ -88,7 +84,7 @@ class ModelState(ModelStateBase):
   inputs: dict[str, np.ndarray]
   prev_desire: np.ndarray
 
-  def __init__(self, cam_w: int, cam_h: int):
+  def __init__(self, cam_w: int, cam_h: int, usbgpu: bool = False):
     ModelStateBase.__init__(self)
 
     env_pkl = os.environ.get('COMBINED_MODEL_PKL')
@@ -103,6 +99,7 @@ class ModelState(ModelStateBase):
     self.LONG_SMOOTH_SECONDS = float(overrides.get('long', ".0"))
     self.MIN_LAT_CONTROL_SPEED = 0.3
     self.PLANPLUS_CONTROL: float = 1.0
+    self.usbgpu = usbgpu
 
     pkl_path = _find_driving_pkl(model_bundle)
     assert pkl_path is not None, "No driving pkl found — all models must be compiled with compile_modeld.py"
@@ -110,24 +107,20 @@ class ModelState(ModelStateBase):
 
   def _init_combined(self, pkl_path, cam_w, cam_h, bundle):
     cloudlog.warning(f"loading combined pkl: {pkl_path}")
-    # TODO-SP: switch to load_oob from openpilot/selfdrive/helpers on next full recompile of all models
-    jits = pickle.load(open_file_chunked(pkl_path))
+    jits = load_oob(open_file_chunked(pkl_path))
 
-    self.DEV = Device.DEFAULT
-    self.WARP_DEV = 'CPU' if USBGPU else self.DEV
+    self.WARP_DEV = 'QCOM' if COMMA_HARDWARE else 'CPU'
+    self.DEV = 'AMD' if self.usbgpu else self.WARP_DEV
     self.QUEUE_DEV = self.DEV
-
     metadata = jits['metadata']
 
-    self._run_policy = jits[(cam_w, cam_h)]['run_policy']
-    self._warp_enqueue = jits[(cam_w, cam_h)]['warp_enqueue']
-
-    # TODO-SP: Remove legacy use_packed detection block after all models are recompiled
-    captured = getattr(self._run_policy, 'captured', None)
-    if captured is not None:
-      use_packed = 'packed_npy_inputs' in getattr(captured, 'expected_names', [])
+    self.is_legacy_model = 'run_policy' not in jits  # remove after next recompile
+    if self.is_legacy_model:
+      self.warp = jits[(cam_w, cam_h)]['warp_enqueue']
+      self.run_policy = jits[(cam_w, cam_h)]['run_policy']
     else:
-      use_packed = True
+      self.run_policy = jits['run_policy']
+      self.warp = jits[(cam_w, cam_h)]
 
     if 'model' in metadata:
       model_metadata = metadata['model']
@@ -136,10 +129,9 @@ class ModelState(ModelStateBase):
       self._policy_slices_list = []
       self._combined_model_type = 'supercombo'
       self._vision_input_names = [key for key in model_metadata['input_shapes'] if 'img' in key]
-      from openpilot.sunnypilot.modeld_v2.compile_modeld import make_supercombo_input_queues
       frame_skip = derive_frame_skip({}, model_metadata['input_shapes'])
       self.input_queues, self.numpy_inputs = make_supercombo_input_queues(model_metadata['input_shapes'],
-                                                                          frame_skip, device=self.QUEUE_DEV, use_packed=use_packed)
+                                                                          frame_skip, device=self.QUEUE_DEV)
     else:
       vision_metadata = metadata['vision']
       policy_keys = [k for k in metadata if k != 'vision']
@@ -152,13 +144,12 @@ class ModelState(ModelStateBase):
       self._policy_slices_list = [metadata[k]['output_slices'] for k in policy_keys]
       self.policy_output_slices = self._policy_slices_list[0]
       self._has_on_policy = any('on' in k.lower() for k in policy_keys)
-      first_policy_metadata = metadata[policy_keys[0]]
-      vision_input_shapes = vision_metadata['input_shapes']
-      policy_input_shapes = first_policy_metadata['input_shapes']
-      self._vision_input_names = [k for k in vision_input_shapes if 'img' in k]
-      frame_skip = derive_frame_skip(vision_input_shapes, policy_input_shapes)
-      self.input_queues, self.numpy_inputs = make_split_input_queues(vision_input_shapes, policy_input_shapes,
-                                                                     frame_skip, device=self.QUEUE_DEV, use_packed=use_packed)
+      self._vision_input_names = [key for key in vision_metadata['input_shapes'] if 'img' in key]
+      first_policy_meta = metadata[policy_keys[0]]
+      frame_skip = derive_frame_skip(vision_metadata['input_shapes'], first_policy_meta['input_shapes'])
+      self.input_queues, self.numpy_inputs = make_split_input_queues(vision_metadata['input_shapes'],
+                                                                     first_policy_meta['input_shapes'],
+                                                                     frame_skip, device=self.QUEUE_DEV)
 
     self._desire_key = next(key for key in self.numpy_inputs if key.startswith('desire'))
     self._road_key = next(key for key in self._vision_input_names if 'big' not in key)
@@ -186,10 +177,33 @@ class ModelState(ModelStateBase):
     self.frame_buf_params = dict.fromkeys(self._vision_input_names, nv12_info)
 
     yuv_size = self.frame_buf_params[self._road_key][3]
-    self._warp_enqueue(
-      **self.input_queues,
-      frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize(),
-      big_frame=Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize())
+    frame_tensor = Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize()
+    big_frame_tensor = Tensor(np.zeros(yuv_size, dtype=np.uint8), device=self.WARP_DEV).contiguous().realize()
+
+    if self.is_legacy_model: # Remove this conditional hack after recompile
+      self.warp(**self.input_queues, frame=frame_tensor, big_frame=big_frame_tensor)
+    else:
+      self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=frame_tensor, big_frame=big_frame_tensor)
+
+    if self.usbgpu:
+      self.warmup()
+
+  def warmup(self) -> None:
+    dummy_frames = {k: np.zeros(self.frame_buf_params[k][3], dtype=np.uint8) for k in self._vision_input_names}
+    transforms = {k: np.eye(3, dtype=np.float32) for k in [self._road_key, self._wide_key] if k}
+
+    dummy_inputs = {}
+    for k, v in self.numpy_inputs.items():
+      if k not in ['tfm', 'big_tfm', 'prev_feat']:
+        dummy_inputs[k] = np.zeros(v.shape, dtype=v.dtype)
+
+    self.run(dummy_frames, transforms, dummy_inputs, prepare_only=False)
+
+    for v in self.numpy_inputs.values():
+      v[:] = 0
+    self.prev_desire[:] = 0
+    self.full_frames.clear()
+    self._blob_cache.clear()
 
 
   @property
@@ -227,11 +241,17 @@ class ModelState(ModelStateBase):
     self.numpy_inputs['tfm'][:, :] = transforms[road_key].reshape(3, 3)
     self.numpy_inputs['big_tfm'][:, :] = transforms[wide_key].reshape(3, 3)
 
-    if prepare_only:
-      self._warp_enqueue(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
-      return None
-
-    raw_outputs = self._run_policy(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+    if self.is_legacy_model:  # remove after next recompile
+      if prepare_only:
+        self.warp(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+        return None
+      raw_outputs = self.run_policy(**self.input_queues, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+    else:
+      if prepare_only:
+        self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+        return None
+      warped = self.warp(**{k: self.input_queues[k] for k in WARP_INPUTS}, frame=self.full_frames[road_key], big_frame=self.full_frames[wide_key])
+      raw_outputs = self.run_policy(**{k: self.input_queues[k] for k in POLICY_INPUTS if k in self.input_queues}, warped=warped)
 
     if self._combined_model_type == 'supercombo':
       model_output = raw_outputs.numpy().flatten()
@@ -267,10 +287,9 @@ class ModelState(ModelStateBase):
       buf[0, :-1] = buf[0, 1:]
       buf[0, -1, :] = outputs['desired_curvature'][0, :] if not self.mlsim else 0
 
-    # TODO-SP: This is a hack to prevent GPU corruption by calculating in CPU space, it can be removed on next recompile
-    if 'prev_feat' not in self.numpy_inputs and 'feat_q' in self.input_queues:
-      feat_val = self.input_queues['feat_q'].numpy()
-      self.input_queues['feat_q'].assign(feat_val).realize()
+    if self.usbgpu and not np.all(np.isfinite(outputs.get('plan', np.array([0.])))):
+      cloudlog.error("model output not finite, dropping frame")
+      return None
 
     return outputs
 
@@ -278,8 +297,8 @@ class ModelState(ModelStateBase):
                             lat_action_t: float, long_action_t: float, v_ego: float) -> log.ModelDataV2.Action:
     if 'action' not in model_output:
       plan = model_output['plan'][0]
-      desired_accel, should_stop = get_accel_from_plan(plan[:, Plan.VELOCITY][:, 0], plan[:, Plan.ACCELERATION][:, 0], self.constants.T_IDXS,
-                                                       action_t=long_action_t)
+      desired_accel = get_accel_from_plan(plan[:, Plan.VELOCITY][:, 0], plan[:, Plan.ACCELERATION][:, 0], self.constants.T_IDXS,
+                                          action_t=long_action_t)
 
       curvature_plan = (plan + (self.PLANPLUS_CONTROL - 1.0) * model_output['planplus'][0]
                         if 'planplus' in model_output and self.PLANPLUS_CONTROL != 1.0 else plan)
@@ -287,8 +306,8 @@ class ModelState(ModelStateBase):
     else:
       desired_accel = model_output['action'][0, 1]
       desired_curvature = model_output['action'][0, 0] / (max(1.0, v_ego))**2
-      should_stop = (v_ego < 0.3 and desired_accel < 0.1)
 
+    stop = v_ego < 0.3 and desired_accel < 0.1
     desired_accel = smooth_value(desired_accel, prev_action.desiredAcceleration, self.LONG_SMOOTH_SECONDS)
 
     if self.generation is not None and self.generation >= 10: # smooth curvature for post FOF models
@@ -297,7 +316,7 @@ class ModelState(ModelStateBase):
       else:
         desired_curvature = prev_action.desiredCurvature
 
-    return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature),desiredAcceleration=float(desired_accel), shouldStop=bool(should_stop))
+    return log.ModelDataV2.Action(desiredCurvature=float(desired_curvature), desiredAcceleration=float(desired_accel), shouldStop=bool(stop))
 
 
 def main(demo=False):
@@ -308,6 +327,14 @@ def main(demo=False):
   setproctitle(PROCESS_NAME)
   config_realtime_process(7, 54)
 
+  USBGPU = usbgpu_present()
+  if USBGPU:
+    os.environ['HCQDEV_WAIT_TIMEOUT_MS'] = '3000'
+
+  params = Params()
+  params.put_bool("UsbGpuLoading", USBGPU)
+  params.remove("UsbGpuActive")
+
   # visionipc clients
   while True:
     available_streams = VisionIpcClient.available_streams("camerad", block=False)
@@ -332,15 +359,34 @@ def main(demo=False):
     cloudlog.warning(f"connected extra cam with buffer size: {vipc_client_extra.buffer_len} ({vipc_client_extra.width} x {vipc_client_extra.height})")
 
   cloudlog.warning("loading model")
-  model = ModelState(cam_w=vipc_client_main.width, cam_h=vipc_client_main.height)
-  cloudlog.warning("models loaded, modeld starting")
+  st = time.monotonic()
+
+  model = None
+  if USBGPU:
+    import threading
+    def load():
+      nonlocal model
+      model = ModelState(cam_w=vipc_client_main.width, cam_h=vipc_client_main.height, usbgpu=True)
+    t = threading.Thread(target=load, daemon=True)
+    t.start()
+    t.join(60)
+    if model is None:
+      params.put_bool("UsbGpuActive", False)
+      raise RuntimeError("eGPU model load failed or timed out (60s)")
+    params.put_bool("UsbGpuActive", True)
+  else:
+    model = ModelState(cam_w=vipc_client_main.width, cam_h=vipc_client_main.height, usbgpu=False)
+
+  params.put_bool("UsbGpuLoading", False)
+  cloudlog.warning(f"models loaded in {time.monotonic() - st:.1f}s, modeld starting")
 
   # messaging
-  pm = PubMaster(["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"])
+  pub_socks = ["modelV2", "drivingModelData", "cameraOdometry", "modelDataV2SP"] + (["chestnutState"] if USBGPU else [])
+  pm = PubMaster(pub_socks)
   sm = SubMaster(["deviceState", "carState", "narrowRoadCameraState", "extrinsicsCalibration", "driverMonitoringState", "carControl", "lateralDelay"])
 
   publish_state = PublishState()
-  params = Params()
+  chestnut_state = ChestnutState(pm, USBGPU) if USBGPU else None
 
   # setup filter to track dropped frames
   frame_dropped_filter = FirstOrderFilter(0., 10., 1. / model.constants.MODEL_FREQ)
@@ -478,6 +524,7 @@ def main(demo=False):
       fill_model_msg(drivingdata_send, modelv2_send, model_output, action,
                      publish_state, meta_main.frame_id, meta_extra.frame_id, frame_id,
                      frame_drop_ratio, meta_main.timestamp_eof, model_execution_time, live_calib_seen, meta_constants)
+      modelv2_send.modelV2.big = model.usbgpu
 
       desire_state = modelv2_send.modelV2.meta.desireState
       l_lane_change_prob = desire_state[log.Desire.laneChangeLeft]
@@ -498,6 +545,8 @@ def main(demo=False):
       pm.send('modelDataV2SP', mdv2sp_send)
     last_vipc_frame_id = meta_main.frame_id
 
+    if chestnut_state is not None and run_count % round(model.constants.MODEL_FREQ / SERVICE_LIST['chestnutState'].frequency) == 0:
+      chestnut_state.send()
 
 if __name__ == "__main__":
   try:
diff --git a/openpilot/sunnypilot/modeld_v2/tests/helpers.py b/openpilot/sunnypilot/modeld_v2/tests/helpers.py
index 6925a61f08..ee59e82785 100644
--- a/openpilot/sunnypilot/modeld_v2/tests/helpers.py
+++ b/openpilot/sunnypilot/modeld_v2/tests/helpers.py
@@ -6,7 +6,6 @@ See the LICENSE.md file in the root directory for more details.
 """
 
 import pathlib
-import pickle
 import tempfile
 
 import openpilot.sunnypilot.models.helpers as helpers
@@ -164,14 +163,16 @@ ARCHETYPES = {
 def make_pkl_data(archetype):
   return {
     'metadata': archetype.metadata_structure,
-    (CAM_W, CAM_H): {'run_policy': _noop_jit, 'warp_enqueue': _noop_jit},
+    'run_policy': _noop_jit,
+    (CAM_W, CAM_H): _noop_jit,
   }
 
 
 def write_pkl(tmp_path, archetype):
+  from openpilot.selfdrive.modeld.helpers import dump_oob
   pkl_path = tmp_path / 'driving_test_tinygrad.pkl'
   with open(pkl_path, 'wb') as f:
-    pickle.dump(make_pkl_data(archetype), f)
+    dump_oob(make_pkl_data(archetype), f)
   return pkl_path
 
 
diff --git a/openpilot/sunnypilot/modeld_v2/tests/test_recovery_power.py b/openpilot/sunnypilot/modeld_v2/tests/test_recovery_power.py
index 85305395ea..fb72022fa5 100644
--- a/openpilot/sunnypilot/modeld_v2/tests/test_recovery_power.py
+++ b/openpilot/sunnypilot/modeld_v2/tests/test_recovery_power.py
@@ -33,7 +33,7 @@ class TestRecoveryPower(OpenpilotTestCase):
 
     def mock_accel(plan_vel, plan_accel, t_idxs, action_t=0.0):
       recorded_vel.append(plan_vel.copy())
-      return 0.0, False
+      return 0.0
 
     def mock_curvature(output, plan, vego, lat_action_t, mlsim):
       recorded_curv_plans.append(plan.copy())
diff --git a/openpilot/sunnypilot/models/fetcher.py b/openpilot/sunnypilot/models/fetcher.py
index eda1117a2a..e60af2925f 100644
--- a/openpilot/sunnypilot/models/fetcher.py
+++ b/openpilot/sunnypilot/models/fetcher.py
@@ -13,6 +13,7 @@ from openpilot.common.params import Params
 from openpilot.common.swaglog import cloudlog
 from openpilot.common.hardware.hw import Paths
 from openpilot.sunnypilot.models.helpers import is_bundle_version_compatible
+from openpilot.selfdrive.modeld.helpers import usbgpu_present
 
 from openpilot.cereal import custom
 
@@ -103,11 +104,11 @@ class ModelParser:
 class ModelCache:
   """Handles caching of model data to avoid frequent remote fetches"""
 
-  def __init__(self, params: Params, cache_timeout: int = int(3600 * 1e9)):
+  def __init__(self, params: Params, cache_timeout: int = int(3600 * 1e9), suffix: str = ""):
     self.params = params
     self.cache_timeout = cache_timeout
-    self._LAST_SYNC_KEY = "ModelManager_LastSyncTime"
-    self._CACHE_KEY = "ModelManager_ModelsCache"
+    self._LAST_SYNC_KEY = f"ModelManager_LastSyncTime{suffix}"
+    self._CACHE_KEY = f"ModelManager_ModelsCache{suffix}"
 
   def _is_expired(self) -> bool:
     """Checks if the cache has expired"""
@@ -139,24 +140,37 @@ 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_v18.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v19.json"
+  MODEL_URL_USBGPU = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_usbgpu_v19.json"
 
   def __init__(self, params: Params):
     self.params = params
-    self.model_cache = ModelCache(params)
     self.model_parser = ModelParser()
+    self._is_usbgpu: bool | None = None
+    self.model_cache = ModelCache(params)
+    self.model_url = self.MODEL_URL
+    self._update_model_source()
+
+  def _update_model_source(self) -> None:
+    """Updates what json to use based on usbgpu availability"""
+    is_usbgpu = usbgpu_present()
+    if is_usbgpu != self._is_usbgpu:
+      self._is_usbgpu = is_usbgpu
+      self.model_cache = ModelCache(self.params, suffix="_USBGPU" if is_usbgpu else "")
+      self.model_url = self.MODEL_URL_USBGPU if is_usbgpu else self.MODEL_URL
+      self.params.put("ModelManager_ActiveJson", self.model_url, block=True)
 
   def _fetch_and_cache_models(self) -> 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.
     """
     try:
-      response = requests.get(self.MODEL_URL, timeout=10)
+      response = requests.get(self.model_url, timeout=10)
 
       # Explicitly handle 404 differently
       if response.status_code == 404:
-        cloudlog.error(f"Models URL returned 404 Not Found: {self.MODEL_URL}")
-        raise HTTPError(f"404 Not Found: {self.MODEL_URL}", response=response)
+        cloudlog.error(f"Models URL returned 404 Not Found: {self.model_url}")
+        raise HTTPError(f"404 Not Found: {self.model_url}", response=response)
 
       # Raise for any other 4xx/5xx
       response.raise_for_status()
@@ -179,6 +193,7 @@ class ModelFetcher:
 
   def get_available_bundles(self) -> list[custom.ModelManagerSP.ModelBundle]:
     """Gets the list of available models, with smart cache handling"""
+    self._update_model_source()
     cached_data, is_expired = self.model_cache.get()
 
     if cached_data and not is_expired:
@@ -202,10 +217,7 @@ if __name__ == "__main__":
   for bundle in bundles:
     for model in bundle.models:
       model_overrides = {override.key: override.value for override in bundle.overrides}
-      # Print model details
       print(f"Bundle: {bundle.internalName}, Type: {model.type}, Status: {bundle.status}, Overrides: {model_overrides}")
-      # Print artifact details
       print(f"Artifact: {model.artifact.fileName}, Download URI: {model.artifact.downloadUri.uri}")
-      # Print metadata details
       if model.artifact.chunks:
         print(f"Contains {len(model.artifact.chunks)} chunks.")
diff --git a/openpilot/sunnypilot/models/helpers.py b/openpilot/sunnypilot/models/helpers.py
index 101d8d196e..b5c97467d3 100644
--- a/openpilot/sunnypilot/models/helpers.py
+++ b/openpilot/sunnypilot/models/helpers.py
@@ -18,7 +18,7 @@ from openpilot.sunnypilot.models.constants import Meta, MetaSimPose, MetaTombRai
 from openpilot.common.hardware.hw import Paths
 
 # SET ME TO THE EXACT JSON VERSION WE SET IN SUNNYPILOT_MODELS REPO
-REQUIRED_JSON_VERSION = 16
+REQUIRED_JSON_VERSION = 17
 
 CUSTOM_MODEL_PATH = Paths.model_root()
 METADATA_PATH = Path(__file__).parent / '../models/supercombo_metadata.pkl'
diff --git a/openpilot/sunnypilot/models/tests/test_tinygrad_ref.py b/openpilot/sunnypilot/models/tests/test_tinygrad_ref.py
index 1712c60410..fd389f93c0 100644
--- a/openpilot/sunnypilot/models/tests/test_tinygrad_ref.py
+++ b/openpilot/sunnypilot/models/tests/test_tinygrad_ref.py
@@ -1,12 +1,13 @@
 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():
-  response = requests.get(ModelFetcher.MODEL_URL, timeout=10)
+  fetcher = ModelFetcher(Params())
+  response = requests.get(fetcher.model_url, timeout=10)
   response.raise_for_status()
   json_data = response.json()
   return json_data.get("tinygrad_ref")
diff --git a/release/ci/model_generator.py b/release/ci/model_generator.py
index 76935f3627..607260f145 100755
--- a/release/ci/model_generator.py
+++ b/release/ci/model_generator.py
@@ -48,6 +48,11 @@ def create_short_name(full_name: str) -> str:
   return result[:8]
 
 
+def create_pkl_name(full_name: str) -> str:
+  pkl = re.sub(r'[^a-zA-Z0-9]+', '_', full_name).strip('_').lower()
+  return pkl
+
+
 def _read_pkl_bytes(pkl_path: Path) -> bytes:
   manifest = Path(f"{pkl_path}.chunkmanifest")
   if manifest.exists():
@@ -154,14 +159,15 @@ if __name__ == "__main__":
   _output_dir = Path(args.output_dir)
   _output_dir.mkdir(exist_ok=True, parents=True)
   _short_name = create_short_name(args.custom_name) if args.custom_name else None
+  _pkl = create_pkl_name(args.custom_name) if args.custom_name else None
 
   _driving_pkl = _find_driving_pkl(_output_dir)
   if not _driving_pkl:
     print(f"No driving_tinygrad.pkl found in {_output_dir}", file=sys.stderr)
     sys.exit(1)
 
-  if _short_name:
-    new_pkl = _output_dir / f"driving_{_short_name.lower()}_tinygrad.pkl"
+  if _pkl:
+    new_pkl = _output_dir / f"driving_{_pkl}_tinygrad.pkl"
     if not new_pkl.exists():
       _driving_pkl = _rename_pkl_with_chunks(_driving_pkl, new_pkl)
     else:
diff --git a/tinygrad_repo b/tinygrad_repo
index 2fecac4e4a..66ee3cfb4f 160000
--- a/tinygrad_repo
+++ b/tinygrad_repo
@@ -1 +1 @@
-Subproject commit 2fecac4e4ac32fe369c41f8400b6e7b9adb18683
+Subproject commit 66ee3cfb4f3a3908a6a20ddfbec7774ba7c09b4e