modeld_v2: emu

* modeld_v2: artifact parity with upstream tg

* m

* m

* Update helpers.py

* ew

* Update convert_upstream_pkl.py

* Update test_convert_upstream_pkl.py

---------

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

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

@@ -34,6 +34,11 @@ on:
         required: false
         type: string
         default: 'qcom'
+      on_device:
+        description: 'Compile on the tici / chestnut runner instead of on a hosted runner with tinygrad''s qcom / amd mock'
+        required: false
+        type: boolean
+        default: false
       hf_repo:
         description: 'Hugging Face dataset repository (e.g. sunnypilot/sunnypilot_models_v1)'
         required: false
@@ -107,6 +112,11 @@ on:
         options:
           - qcom
           - chestnut
+      on_device:
+        description: 'Compile on the tici / chestnut runner instead of on a hosted runner with tinygrad''s qcom / amd mock'
+        required: false
+        type: boolean
+        default: false
       hf_repo:
         description: 'Hugging Face dataset repository'
         required: false
@@ -130,11 +140,45 @@ jobs:
       is_20hz: ${{ inputs.is_20hz }}
       artifact_suffix: ${{ inputs.artifact_suffix }}
       target_hardware: ${{ inputs.target_hardware }}
+      on_device: ${{ inputs.on_device || false }}
     secrets: inherit
 
-  publish_model:
+  mock_parity:
     if: ${{ !cancelled() && needs.build_model.result == 'success' }}
     needs: build_model
+    runs-on: ubuntu-24.04
+    permissions:
+      actions: read
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - uses: ./.github/workflows/setup-amd-mock
+        if: inputs.target_hardware == 'chestnut'
+      - run: python3 -m pip install -q numpy ${{ inputs.target_hardware != 'chestnut' && 'tinymesa==25.2.7.2' || '' }}
+      - name: Download artifact name file
+        uses: actions/download-artifact@v4
+        with:
+          name: artifact-name-${{ inputs.custom_name || inputs.upstream_branch }}
+          path: artifact_name
+      - name: Download the built model
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          ARTIFACT_NAME=$(cat artifact_name/artifact_name.txt)
+          id=$(gh api -X GET "repos/$GITHUB_REPOSITORY/actions/runs/$GITHUB_RUN_ID/artifacts" -f name="$ARTIFACT_NAME" --jq '.artifacts[0].id')
+          url=$(curl -sf -o /dev/null -w '%{redirect_url}' -H "Authorization: Bearer $GH_TOKEN" "https://api.github.com/repos/$GITHUB_REPOSITORY/actions/artifacts/$id/zip")
+          sudo apt-get install -y -qq aria2 > /dev/null
+          aria2c -q -x16 -s16 -k8M --max-tries=5 --file-allocation=none -o model.zip "$url" && unzip -q model.zip -d output && rm model.zip
+      - name: Replay its check frame on the mock and require the hardware's result bit for bit
+        env:
+          PYTHONPATH: ${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}
+          DEV: ${{ inputs.target_hardware == 'chestnut' && 'MOCKUSB+AMD' || 'MOCK+QCOM' }}
+        run: python3 -m openpilot.sunnypilot.modeld_v2.mock_parity check output/${{ inputs.target_hardware == 'chestnut' && 'big_' || '' }}driving_tinygrad.pkl
+
+  publish_model:
+    if: ${{ !cancelled() && needs.build_model.result == 'success' && contains(fromJSON('["success", "skipped"]'), needs.mock_parity.result) }}
+    needs: [build_model, mock_parity]
     runs-on: ubuntu-latest
     permissions:
       actions: read

.github/workflows/chestnut-build-parity/action.yml

@@ -0,0 +1,33 @@
+name: Chestnut mock build parity
+description: >-
+  Convert one commaai precompiled chestnut model (LS2) on this runner under tinygrad's amd mock, as a hosted chestnut build does, and require the
+  exact bytes the chestnut runner built (the published artifact). Needs setup-amd-mock, numpy and zstandard; tinygrad_repo checked out.
+runs:
+  using: composite
+  steps:
+    - name: Rebuild LS2 on the mock and require the chestnut runner's artifact byte for byte
+      shell: bash
+      env:
+        REF: b92fc4828664f8767bce5e92f1630ed319cbe509  # commaai/openpilot: precompiled big_driving_tinygrad.pkl, pins tinygrad ec91b74b5b
+        PKL_OID: b4dead3ee091d66ab002a6b8e716c26f4f79e3c931e4be5f48e017ba92f6e960
+        PKL_SIZE: "767228907"
+        TG_SHA: ec91b74b5b6b0fc924b72d3f27689679ca2438b9
+        WANT: 680a8d629785f2b1d1a2b0e40d0dfbcd0f7291c21e20c51259f1d768a22ec217  # driving_light_speed_2_november_04_2026_tinygrad.pkl, recompiled46
+      run: |
+        set -euo pipefail
+        work="$RUNNER_TEMP/chestnut_parity" && rm -rf "$work" && mkdir -p "$work"
+        href=$(curl -fsSL --retry 3 -X POST -H "Accept: application/vnd.git-lfs+json" -H "Content-Type: application/vnd.git-lfs+json" \
+          https://huggingface.co/commaai/openpilot-lfs.git/info/lfs/objects/batch \
+          -d "{\"operation\":\"download\",\"transfer\":[\"basic\"],\"objects\":[{\"oid\":\"$PKL_OID\",\"size\":$PKL_SIZE}]}" | jq -r '.objects[0].actions.download.href')
+        curl -fsSL --retry 3 -o "$work/big_driving_tinygrad.pkl" "$href"
+        echo "$PKL_OID  $work/big_driving_tinygrad.pkl" | sha256sum -c -
+        tg="$work/tinygrad" && git init -q "$tg" && git -C "$tg" fetch -q --depth 1 https://github.com/tinygrad/tinygrad "$TG_SHA" && git -C "$tg" checkout -q FETCH_HEAD
+        rm -rf "$tg/test/mockgpu" && cp -r "$GITHUB_WORKSPACE/tinygrad_repo/test/mockgpu" "$tg/test/mockgpu"
+        LLVM_PATH="${LLVM_PATH:-$(ls /usr/lib/llvm-20/lib/libLLVM.so* | head -1)}" DEV=MOCKUSB+AMD CHECK_OOB=0 GMMU=0 PYTHONPATH="$tg:$GITHUB_WORKSPACE" \
+          python3 openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py "$work/big_driving_tinygrad.pkl" "$work/LS2.pkl"
+        got=$(sha256sum "$work/LS2.pkl" | cut -d' ' -f1)
+        if [ "$got" != "$WANT" ]; then
+          echo "::error::the mock's LS2 build ($got) is not the chestnut runner's ($WANT): this runner's LLVM or the mock no longer builds what the chestnut does"
+          exit 1
+        fi
+        echo "LS2 rebuilt on the mock: byte-identical to the chestnut runner's build"

.github/workflows/setup-amd-mock/action.yml

@@ -0,0 +1,18 @@
+name: Set up tinygrad's amd mock
+description: LLVM 20 and ROCm comgr, as tinygrad_repo/.github/actions/setup-tinygrad installs them for the mock gpu
+runs:
+  using: composite
+  steps:
+    - name: Install LLVM 20 and ROCm comgr
+      shell: bash
+      run: |
+        . /etc/os-release
+        curl -fsSL https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc > /dev/null
+        echo "deb http://apt.llvm.org/$VERSION_CODENAME/ llvm-toolchain-$VERSION_CODENAME-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list
+        sudo mkdir -p /etc/apt/keyrings
+        curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+        echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $VERSION_CODENAME main" | sudo tee /etc/apt/sources.list.d/rocm.list
+        printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' | sudo tee /etc/apt/preferences.d/rocm-pin-600
+        sudo apt-get -qq update && sudo apt-get -qq install -y --no-install-recommends libllvm20 clang-20 lld-20 comgr
+        printf '/opt/rocm/lib\n/opt/rocm/lib64\n' | sudo tee /etc/ld.so.conf.d/rocm.conf && sudo ldconfig
+        echo "/usr/lib/llvm-20/bin" >> "$GITHUB_PATH"

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

@@ -35,6 +35,11 @@ on:
         required: false
         type: string
         default: 'qcom'
+      on_device:
+        description: 'Compile on the tici / chestnut runner instead of on a hosted runner with tinygrad''s qcom / amd mock'
+        required: false
+        type: boolean
+        default: false
   workflow_dispatch:
     inputs:
       upstream_branch:
@@ -59,6 +64,11 @@ on:
           - qcom
           - chestnut
         default: 'qcom'
+      on_device:
+        description: 'Compile on the tici / chestnut runner instead of on a hosted runner with tinygrad''s qcom / amd mock'
+        required: false
+        type: boolean
+        default: false
 
 
 run-name: Build model [${{ inputs.custom_name || inputs.upstream_branch }}] from ref [${{ inputs.upstream_branch }}]
@@ -153,18 +163,48 @@ jobs:
           if-no-files-found: error
 
   build_model:
-    runs-on: [self-hosted, "${{ inputs.target_hardware == 'chestnut' && 'chestnut' || 'tici' }}"]
+    runs-on: ${{ inputs.on_device && fromJSON(inputs.target_hardware == 'chestnut' && '["self-hosted","chestnut"]' || '["self-hosted","tici"]') || 'ubuntu-24.04' }}
     needs: get_model
     env:
       MODEL_NAME: ${{ inputs.custom_name || inputs.upstream_branch }} (${{ needs.get_model.outputs.model_date }})
       REF: ${{ inputs.upstream_branch }}
+      MOCK_BUILD: ${{ !inputs.on_device }}
     steps:
       - uses: actions/checkout@v4
         with:
           fetch-depth: 1
           submodules: recursive
 
+      - name: Set up the hosted mock build
+        if: env.MOCK_BUILD == 'true'
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+
+      - name: Install LLVM 20 and ROCm comgr for the amd mock
+        if: env.MOCK_BUILD == 'true' && inputs.target_hardware == 'chestnut'
+        uses: ./.github/workflows/setup-amd-mock
+
+      - name: Install the amd mock build dependencies
+        if: env.MOCK_BUILD == 'true' && inputs.target_hardware == 'chestnut'
+        run: |
+          python3 -m pip install -q numpy zstandard
+          echo "LLVM_PATH=$(ls /usr/lib/llvm-20/lib/libLLVM.so* | head -1)" >> "$GITHUB_ENV"  # the chestnut runner's LLVM, never a newer one
+
+      - name: Check this runner builds what the chestnut runner does
+        if: env.MOCK_BUILD == 'true' && inputs.target_hardware == 'chestnut'
+        uses: ./.github/workflows/chestnut-build-parity
+
+      - name: Install the qcom mock build dependencies and check the compiler
+        if: env.MOCK_BUILD == 'true' && inputs.target_hardware != 'chestnut'
+        run: |
+          sudo apt-get -qq update && sudo apt-get -qq install -y --no-install-recommends qemu-user-static aria2
+          python3 -m pip install -q numpy zstandard tinymesa==25.2.7.2
+          # every kernel an a630 has compiled must compile to the same bytes here before a model is built here
+          PYTHONPATH="$TINYGRAD_PATH" python3 "$TINYGRAD_PATH/test/device/qcom/compiler_parity.py" check "$TINYGRAD_PATH/test/mockgpu/qcom/a630_compiled.pkl.xz"
+
       - name: Set environment variables
+        if: env.MOCK_BUILD != 'true'
         id: set-env
         run: |
           # Set up common environment
@@ -177,6 +217,7 @@ jobs:
           printenv >> $GITHUB_ENV
 
       - name: Setup build environment
+        if: env.MOCK_BUILD != 'true'
         run: |
           mkdir -p "${BUILD_DIR}/"
           sudo find $BUILD_DIR/ -mindepth 1 -delete
@@ -198,21 +239,28 @@ jobs:
 
       - name: Build Model
         run: |
-          source /etc/profile
-          export UV_PROJECT_ENVIRONMENT=${HOME}/venv
-          export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
-          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+          if [ "$MOCK_BUILD" = "true" ]; then
+            TASKSET=""
+            export CHECK_OOB=0 GMMU=0  # as test/mockgpu runs the amd mock
+          else
+            source /etc/profile
+            export UV_PROJECT_ENVIRONMENT=${HOME}/venv
+            export VIRTUAL_ENV=$UV_PROJECT_ENVIRONMENT
+            source ${UV_PROJECT_ENVIRONMENT}/bin/activate
+            TASKSET="taskset -c 7"
+          fi
           export PYTHONPATH="${PYTHONPATH}:${{ env.TINYGRAD_PATH }}:${{ github.workspace }}"
 
           COMPILE_MODELD="${{ github.workspace }}/openpilot/sunnypilot/modeld_v2/compile_modeld.py"
           MODEL_SIZE=$(python3 -c "from openpilot.common.transformations.model import MEDMODEL_INPUT_SIZE as s; print(f'{s[0]}x{s[1]}')")
           CAMERA_RES=$(python3 -c "from openpilot.common.transformations.camera import _ar_ox_fisheye as a, _os_fisheye as o; print(f'{a.width}x{a.height} {o.width}x{o.height}')")
 
-          TG_FLAGS_QCOM="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 OPENPILOT_HACKS=1"
+          MOCK=$([ "$MOCK_BUILD" = "true" ] && echo MOCK || true)
+          TG_FLAGS_QCOM="DEV=${MOCK}QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 OPENPILOT_HACKS=1"
           if [ "${{ inputs.target_hardware }}" == "chestnut" ]; then
             echo "CHESTNUT build"
             export CHESTNUT=1
-            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32"
+            TG_FLAGS="DEBUG=1 DEV=${MOCK}USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32"
             OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
           else
             echo "QCOM build"
@@ -227,14 +275,17 @@ jobs:
             echo "Converting commaai's precompiled model with tinygrad $TG_SHA"
             UPSTREAM_TG="${{ runner.temp }}/upstream_tinygrad"
             rm -rf "$UPSTREAM_TG" && git init -q "$UPSTREAM_TG" && git -C "$UPSTREAM_TG" fetch -q --depth 1 "$TG_URL" "$TG_SHA" && git -C "$UPSTREAM_TG" checkout -q FETCH_HEAD
+            if [ "$MOCK_BUILD" = "true" ]; then  # convert under commaai's tinygrad, on our mock: it is the one measured against a chestnut
+              rm -rf "$UPSTREAM_TG/test/mockgpu" && cp -r "$TINYGRAD_PATH/test/mockgpu" "$UPSTREAM_TG/test/mockgpu"
+            fi
             CONVERTED="${{ runner.temp }}/big_driving_tinygrad_plain.pkl"
-            env DEV=USB+AMD PYTHONPATH="$UPSTREAM_TG:${{ github.workspace }}" taskset -c 7 python3 openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py \
+            env DEV=${MOCK}USB+AMD PYTHONPATH="$UPSTREAM_TG:${{ github.workspace }}" $TASKSET python3 openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py \
               "${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl" "$CONVERTED"
             mv "$CONVERTED" "$OUTPUT_PKL"
           elif NEW_ARCH_ONNX=$(find "${{ env.MODELS_DIR }}" -maxdepth 1 -name "${{ inputs.target_hardware == 'chestnut' && 'big_' || '' }}*driving*.onnx" -print -quit) &&
                [ -n "$NEW_ARCH_ONNX" ] && grep -qa new_img "$NEW_ARCH_ONNX"; then # the new architecture: one ONNX taking the warped frames as new_img
             echo "Compiling new architecture ONNX: $NEW_ARCH_ONNX"
-            env ${TG_FLAGS} taskset -c 7 python3 openpilot/sunnypilot/modeld_v2/compile_modeld_v2.py \
+            env ${TG_FLAGS} $TASKSET python3 openpilot/sunnypilot/modeld_v2/compile_modeld_v2.py \
               "$NEW_ARCH_ONNX" "$OUTPUT_PKL" --device-input "*" --benchmark-runs 1
           else
             echo "Detected onnx architecture. Routing to native compile_modeld.py."
@@ -284,7 +335,7 @@ jobs:
 
             if [ -n "$MODEL_TYPE" ]; then
               echo "Detected: $MODEL_TYPE -> $OUTPUT_PKL"
-              env ${TG_FLAGS} taskset -c 7 python3 "$COMPILE_MODELD" \
+              env ${TG_FLAGS} $TASKSET python3 "$COMPILE_MODELD" \
                 --model-type $MODEL_TYPE \
                 --model-size $MODEL_SIZE \
                 --camera-resolutions $CAMERA_RES \
@@ -304,7 +355,7 @@ jobs:
             --include='/${{ inputs.target_hardware == 'chestnut' && 'big_' || '' }}driving_tinygrad.pkl' \
             --exclude='*' \
             --delete-excluded \
-            --chown=comma:comma \
+            $([ "$MOCK_BUILD" = "true" ] || echo --chown=comma:comma) \
             ${{ env.MODELS_DIR }}/ ${{ env.OUTPUT_DIR }}/
 
           python3 "${{ github.workspace }}/release/ci/model_generator.py" \
@@ -333,7 +384,7 @@ jobs:
           path: ${{ env.OUTPUT_DIR }}/artifact_name.txt
 
       - name: Re-enable powersave
-        if: always()
+        if: always() && env.MOCK_BUILD != 'true'
         run: |
           source ${UV_PROJECT_ENVIRONMENT}/bin/activate
           PYTHONPATH=$PYTHONPATH:${{ github.workspace }}/ ${{ github.workspace }}/scripts/manage-powersave.py --enable

.github/workflows/test_models.yml

@@ -28,24 +28,30 @@ jobs:
         env:
           DEV: CPU
         run: tools/op.sh test -v openpilot/sunnypilot/modeld_v2/tests/test_plain_artifact.py
-      - name: Install LLVM 20 and ROCm comgr for the amd mock, as tinygrad_repo/.github/actions/setup-tinygrad does
-        run: |
-          . /etc/os-release
-          curl -fsSL https://apt.llvm.org/llvm-snapshot.gpg.key | sudo tee /etc/apt/trusted.gpg.d/apt.llvm.org.asc > /dev/null
-          echo "deb http://apt.llvm.org/$VERSION_CODENAME/ llvm-toolchain-$VERSION_CODENAME-20 main" | sudo tee /etc/apt/sources.list.d/llvm.list
-          sudo mkdir -p /etc/apt/keyrings
-          curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
-          echo "deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/7.1 $VERSION_CODENAME main" | sudo tee /etc/apt/sources.list.d/rocm.list
-          printf 'Package: *\nPin: release o=repo.radeon.com\nPin-Priority: 600\n' | sudo tee /etc/apt/preferences.d/rocm-pin-600
-          sudo apt-get -qq update && sudo apt-get -qq install -y --no-install-recommends libllvm20 clang-20 lld-20 comgr
-          printf '/opt/rocm/lib\n/opt/rocm/lib64\n' | sudo tee /etc/ld.so.conf.d/rocm.conf && sudo ldconfig
-          echo "/usr/lib/llvm-20/bin" >> "$GITHUB_PATH"
+      - uses: ./.github/workflows/setup-amd-mock
       - name: Convert a pickled USB+AMD jit and replay it under tinygrad's amd mock
         env:
           DEV: MOCKUSB+AMD
           CHECK_OOB: 0
           GMMU: 0
         run: tools/op.sh test -v openpilot/sunnypilot/modeld_v2/tests/test_convert_upstream_pkl.py
+      - name: Resolve commaai's tinygrad pin, which their precompiled chestnut model is built and converted with
+        id: upstream-tg
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          read -r sha url < <(gh api repos/commaai/openpilot/contents/tinygrad_repo --jq '.sha + " " + .submodule_git_url')
+          echo "commaai/openpilot master pins $url@$sha, ours is $(git rev-parse HEAD:tinygrad_repo)"
+          dir="$RUNNER_TEMP/upstream_tinygrad"
+          git init -q "$dir" && git -C "$dir" fetch -q --depth 1 "$url" "$sha" && git -C "$dir" checkout -q FETCH_HEAD
+          echo "path=$dir" >> "$GITHUB_OUTPUT"
+      - name: Convert a pickled USB+AMD jit under commaai's tinygrad, as the chestnut model build does
+        env:
+          DEV: MOCKUSB+AMD
+          CHECK_OOB: 0
+          GMMU: 0
+          PYTHONPATH: ${{ steps.upstream-tg.outputs.path }}:${{ github.workspace }}
+        run: tools/op.sh test -v openpilot/sunnypilot/modeld_v2/tests/test_convert_upstream_pkl.py
       - name: Where tinygrad moved, for the helper or failure point named above
         if: failure() && github.event_name == 'pull_request'
         run: |
@@ -63,6 +69,8 @@ jobs:
     runs-on: ubuntu-latest
     outputs:
       models: ${{ steps.set-matrix.outputs.models }}
+      chestnut: ${{ steps.set-matrix.outputs.chestnut }}
+      qcom: ${{ steps.set-matrix.outputs.qcom }}
     steps:
       - uses: actions/checkout@v4
       - name: Fetch and Parse json
@@ -74,16 +82,22 @@ jobs:
           with open("openpilot/sunnypilot/models/fetcher.py", "r") as f:
             urls = re.findall(r"MODEL_URL(?:_CHESTNUT)?\s*=\s*[\"'"'"']([^\"'"'"']+)[\"'"'"']", f.read())
 
-          artifacts = []
+          artifacts, chestnut, qcom = [], [], []
           for url in urls:
             data = json.loads(urllib.request.urlopen(url).read())
             for bundle in data.get("bundles", []):
               for model in bundle.get("models", []):
                 if "artifact" in model:
                   artifacts.append(model["artifact"])
+                  if "chestnut" in url:
+                    chestnut.append(model["artifact"])
+                  elif "tinygrad" in model["artifact"]["file_name"]:
+                    qcom.append(model["artifact"])
 
           with open(os.environ["GITHUB_OUTPUT"], "a") as f:
             f.write(f"models={json.dumps(artifacts)}\n")
+            f.write(f"chestnut={json.dumps(chestnut)}\n")
+            f.write(f"qcom={json.dumps(qcom)}\n")
           '
 
   test-model:
@@ -120,3 +134,74 @@ jobs:
           export PYTHONPATH="$GITHUB_WORKSPACE/tinygrad_repo:$GITHUB_WORKSPACE:$PYTHONPATH"
           python3 -m pip uninstall -y tinygrad || uv pip uninstall -y tinygrad --system || true
           python3 -m pytest openpilot/sunnypilot/modeld_v2/tests/test_models.py
+
+  chestnut-mock-parity:
+    name: Mock parity ${{ matrix.artifact.file_name }}
+    needs: generate-matrix
+    runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        artifact: ${{ fromJson(needs.generate-matrix.outputs.chestnut) }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - uses: ./.github/workflows/setup-amd-mock
+      - run: python3 -m pip install -q numpy
+      - name: Download model
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          MODEL_URL: ${{ matrix.artifact.download_uri.url }}
+          MODEL_SHA: ${{ matrix.artifact.download_uri.sha256 }}
+        shell: bash
+        run: |
+          AUTH=()
+          if [ -n "$HF_TOKEN" ]; then AUTH=(-H "Authorization: Bearer $HF_TOKEN"); fi
+          curl --retry 3 --retry-all-errors -s -S -f -L "${AUTH[@]}" -o "$RUNNER_TEMP/model.pkl" "$MODEL_URL"
+          if [ -n "$MODEL_SHA" ]; then echo "$MODEL_SHA  $RUNNER_TEMP/model.pkl" | sha256sum -c -; fi
+      - name: Replay its check frame on the amd mock and require the chestnut's result bit for bit
+        env:
+          PYTHONPATH: ${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}
+        run: python3 -m openpilot.sunnypilot.modeld_v2.mock_parity check "$RUNNER_TEMP/model.pkl"
+
+  qcom-mock-parity:
+    name: QCOM mock parity ${{ matrix.artifact.file_name }}
+    needs: generate-matrix
+    runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        artifact: ${{ fromJson(needs.generate-matrix.outputs.qcom) }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - run: python3 -m pip install -q numpy tinymesa==25.2.7.2
+      - name: Download model
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          MODEL_URL: ${{ matrix.artifact.download_uri.url }}
+          MODEL_SHA: ${{ matrix.artifact.download_uri.sha256 }}
+        shell: bash
+        run: |
+          AUTH=()
+          if [ -n "$HF_TOKEN" ]; then AUTH=(-H "Authorization: Bearer $HF_TOKEN"); fi
+          curl --retry 3 --retry-all-errors -s -S -f -L "${AUTH[@]}" -o "$RUNNER_TEMP/model.pkl" "$MODEL_URL"
+          if [ -n "$MODEL_SHA" ]; then echo "$MODEL_SHA  $RUNNER_TEMP/model.pkl" | sha256sum -c -; fi
+      - name: Replay its check frame on the qcom mock and require the comma device's result bit for bit
+        env:
+          DEV: MOCK+QCOM
+          PYTHONPATH: ${{ github.workspace }}/tinygrad_repo:${{ github.workspace }}
+        run: python3 -m openpilot.sunnypilot.modeld_v2.mock_parity check "$RUNNER_TEMP/model.pkl"
+
+  chestnut-mock-build-parity:
+    name: Chestnut build on the amd mock = chestnut runner build
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: true
+      - uses: ./.github/workflows/setup-amd-mock
+      - run: python3 -m pip install -q numpy zstandard
+      - uses: ./.github/workflows/chestnut-build-parity

openpilot/sunnypilot/modeld_v2/compile_modeld.py

@@ -371,7 +371,7 @@ def _compile_warp_resolution_worker(cam_w, cam_h, model_w, model_h, benchmark_ru
 
 
 if __name__ == "__main__":
-  if 'USB' in os.getenv('DEV', '') or os.getenv('CHESTNUT'):
+  if ('USB' in os.getenv('DEV', '') or os.getenv('CHESTNUT')) and not os.getenv('DEV', '').startswith('MOCK'):
     from openpilot.system.hardware.chestnut.flash import link_up
     for _ in range(10):
       if link_up():

openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py

@@ -66,47 +66,75 @@ def _patched(buf, topo) -> tuple[bytearray, dict]:
   return blob, rows
 
 
-def _evaluate(u, env):
+def _packed_args(program) -> dict[int, int]:
+  from tinygrad.device import TinyELF
+  params = getattr(program.arg, "params", None) or ()
+  return {off: p.arg.slot for (off, _), p in zip(TinyELF.iter_sig(tuple((None, p.addrspace, p.dtype, ()) for p in params)), params, strict=True)}
+
+
+def _outer_slot(u, packed: dict[int, int]) -> int | None:
+  from tinygrad.uop.ops import Ops
+  if u.op is Ops.PARAM:
+    return None if packed else u.arg.slot
+  if u.op.name == "REFPARAM":
+    return u.arg.slot
+  if packed and u.op is Ops.LOAD and u.src[0].op is Ops.INDEX and _root(u.src[0].src[0]).op is Ops.PARAM:
+    index = _evaluate(u.src[0].src[1], {}, packed)
+    return None if index is None else packed.get(index * u.dtype.itemsize)
+  return None
+
+
+def _evaluate(u, env, packed: dict[int, int]):
   from tinygrad.uop.ops import GroupOp, Ops, exec_alu
   if u.op in (Ops.AFTER, Ops.CAST, Ops.BITCAST):
-    return _evaluate(u.src[0], env)
+    return _evaluate(u.src[0], env, packed)
   if u.op is Ops.CONST:
     return u.arg
-  if u.op in (Ops.PARAM, Ops.RANGE):
-    return env.get(u.arg.slot if u.op is Ops.PARAM else u)
+  if u.op is Ops.RANGE:
+    return env.get(u)
+  if (slot := _outer_slot(u, packed)) is not None:
+    return env.get(slot)
   if u.op is Ops.LOAD:
-    table, index = _root(u.src[0].src[0]), _evaluate(u.src[0].src[1], env)
+    table, index = _root(u.src[0].src[0]), _evaluate(u.src[0].src[1], env, packed)
     return struct.unpack_from(u.dtype.fmt, table.arg, index * u.dtype.itemsize)[0] if table.op is Ops.BINARY and index is not None else None
-  if u.op in GroupOp.ALU and None not in (values := [_evaluate(x, env) for x in u.src]):
+  if u.op in GroupOp.ALU and None not in (values := [_evaluate(x, env, packed) for x in u.src]):
     return exec_alu(u.op, u.dtype, values, *([False] if "truncate_output" in inspect.signature(exec_alu).parameters else []))
   return None
 
 
 def _input_pokes(submit, kernargs) -> dict[int, tuple[int, int]]:
   from tinygrad.uop.ops import Ops
-  outer = dict(enumerate(submit.src[1:]))
+  outer, packed = dict(enumerate(submit.src[1:])), _packed_args(submit.src[0])
   kernargs_params = {i for i, a in outer.items() if a.op is Ops.GETADDR and _root(a.src[0]) is kernargs}
   inputs = {i: _root(a.src[0]).arg.slot for i, a in outer.items() if a.op is Ops.GETADDR and _root(a.src[0]).op is Ops.PARAM}
   pokes = {}
   for call in (u for u in submit.src[0].src[0].toposort() if u.op is Ops.CALL and len(u.src) > 3):
     _, address, value, *_ = call.src[1:]
-    used = list({x.arg.slot for x in value.toposort() if x.op is Ops.PARAM} & inputs.keys())
-    if len(used) != 1 or not kernargs_params & {x.arg.slot for x in address.toposort() if x.op is Ops.PARAM}:
+    used = list({slot for x in value.toposort() if (slot := _outer_slot(x, packed)) is not None} & inputs.keys())
+    if len(used) != 1 or not kernargs_params & {_outer_slot(x, packed) for x in address.toposort()}:
       continue
     ranges = list(dict.fromkeys(x for x in (*address.toposort(), *value.toposort()) if x.op is Ops.RANGE))
     for trip in itertools.product(*[range(int(r.vmax) + 1) for r in ranges]):
       env = dict.fromkeys(outer, 0) | dict(zip(ranges, trip, strict=True))
-      if (offset := _evaluate(address, env)) % 8 == 0:
-        pokes[offset] = (inputs[used[0]], _evaluate(value, env))
+      if (offset := _evaluate(address, env, packed)) % 8 == 0:
+        pokes[offset] = (inputs[used[0]], _evaluate(value, env, packed))
   return pokes
 
 
-def _elf(image: bytes, rodata: int) -> bytes:
-  names, offset, data, headers = b"\0.text\0.rodata\0.shstrtab\0", 64, b"", []
-  for name, kind, address, content in [(0, 0, 0, b""), (1, 1, ALIGNMENT, image[:rodata]), (7, 1, ALIGNMENT + rodata, image[rodata:]), (15, 3, 0, names)]:
-    headers.append(struct.pack("<IIQQQQIIQQ", name, kind, 2 if kind == 1 else 0, address, offset, len(content), 0, 0, 1, 0))
+def _elf(image: bytes, rodata: int, entry: int) -> bytes:
+  spans = [(".text", 0, rodata), (".rodata", rodata, len(image))] if entry < rodata else \
+          [(".data", 0, rodata), (".rodata", rodata, entry), (".text", entry, len(image))]
+  sections, names = [(0, 0, 0, 0, b"")], b"\0"
+  for name, start, end in (span for span in spans if span[1] < span[2]):
+    sections.append((len(names), 1, 6 if name == ".text" else 2, ALIGNMENT + start, image[start:end]))
+    names += name.encode() + b"\0"
+  sections.append((len(names), 3, 0, 0, names + b".shstrtab\0"))
+  offset, data, headers = 64, b"", []
+  for name, kind, flags, address, content in sections:
+    headers.append(struct.pack("<IIQQQQIIQQ", name, kind, flags, address, offset, len(content), 0, 0, 1, 0))
     data, offset = data + content, offset + len(content)
-  return b"\x7fELF\x02\x01\x01" + bytes(9) + struct.pack("<HHIQQQIHHHHHH", 1, 0xe0, 1, 0, 0, offset, 0, 64, 0, 0, 64, 4, 3) + data + b"".join(headers)
+  header = struct.pack("<HHIQQQIHHHHHH", 1, 0xe0, 1, 0, 0, offset, 0, 64, 0, 0, 64, len(sections), len(sections) - 1)
+  return b"\x7fELF\x02\x01\x01" + bytes(9) + header + data + b"".join(headers)
 
 
 def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
@@ -155,7 +183,7 @@ def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
     if image_alloc not in programs:
       image = bytes(_patched(image_alloc, topo)[0])
       rodata = next(r for r in range(0, len(image) - 24, 64) if r + struct.unpack_from("<q", image, r + 16)[0] == entry)
-      programs[image_alloc] = blob.add(_elf(image, rodata))
+      programs[image_alloc] = blob.add(_elf(image, rodata, entry))
     start, size = _view(arguments.src[0])
     args = []
     for slot in range(start, start + unwrap(size), 8):
@@ -166,7 +194,8 @@ def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
         base, at, nbytes = _address(argument_words[slot])
         args.append({"buffer": buffers.setdefault(base, len(buffers)), "offset": at, "nbytes": nbytes or base.max_numel() * base.dtype.itemsize, "dtype": u8})
     outs, ins = kernel[6]
-    calls.append({"kind": "program", "name": re.sub(r"\x1b\[[0-9;]*m", "", kernel[1]), "lib": programs[image_alloc], "target": dataclasses.asdict(target),
+    device_target = dataclasses.asdict(dataclasses.replace(target, interface=target.interface.removeprefix("MOCK")))
+    calls.append({"kind": "program", "name": re.sub(r"\x1b\[[0-9;]*m", "", kernel[1]), "lib": programs[image_alloc], "target": device_target,
                   "signature": [[f"data{i}", i, u8, [a.get("nbytes", 1)]] for i, a in enumerate(args)], "global_size": grid, "local_size": local,
                   "globals": list(range(len(args))), "outs": list(outs), "ins": list(ins), "args": args})
     needs_init |= {args[i]["buffer"] for i in ins if "buffer" in args[i] and args[i]["buffer"] not in written}

openpilot/sunnypilot/modeld_v2/helpers.py

@@ -264,8 +264,8 @@ def _program_uop(call: dict, lib: bytes):
     parameter = UOp.param(call["globals"][compact_slot], _dtype(dtype), math.prod(shape))
     image = (shape[0], shape[1]) if len(shape) == 3 and shape[2] == 4 else None
     parameters.append(parameter.replace(arg=_fields(parameter.arg, name=name, image=image)))
-  info = _fields(ProgramInfo, global_size=tuple(call["global_size"]), local_size=tuple(call["local_size"]), globals=tuple(call["globals"]),
-                 outs=tuple(call["outs"]), ins=tuple(call["ins"]), target=_fields(Target, **call["target"]))
+  info = _fields(ProgramInfo, global_size=tuple(call["global_size"]), local_size=tuple(call["local_size"]), params=tuple(parameters),
+                 globals=tuple(call["globals"]), outs=tuple(call["outs"]), ins=tuple(call["ins"]), target=_fields(Target, **call["target"]))
   return UOp(Ops.PROGRAM, src=(UOp(Ops.SINK, arg=KernelInfo(name=call["name"])), UOp(Ops.LINEAR, src=tuple(parameters)),
                                UOp(Ops.SOURCE, arg=""), UOp(Ops.BINARY, arg=lib)), arg=info)
 

openpilot/sunnypilot/modeld_v2/mock_parity.py

@@ -0,0 +1,152 @@
+import json
+import os
+import shutil
+import struct
+import sys
+import tempfile
+import time
+
+os.environ.setdefault("DEV", "MOCKUSB+AMD")
+QCOM = "QCOM" in os.environ["DEV"]
+ON_MOCK = os.environ["DEV"].startswith("MOCK")
+ARCH, HARDWARE = ("a630", "comma 3x/4 (a630)") if QCOM else ("gfx1200", "chestnut")
+os.environ.setdefault("CHECK_OOB", "0")
+os.environ.setdefault("GMMU", "0")
+if sys.platform == "darwin" and os.environ["DEV"].startswith("MOCKUSB"):  # the mock's VRAM is a memfd, which macOS lacks: an unlinked file is the same
+  from tinygrad.runtime.autogen import libc
+
+  def _memfd_create(name, flags):
+    fd, path = tempfile.mkstemp()
+    os.unlink(path)
+    return fd
+  libc.memfd_create = _memfd_create
+
+import numpy as np
+
+from openpilot.sunnypilot.modeld_v2.helpers import ALIGNMENT, MAGIC, PlainJit, _device, read_parts
+
+
+class ParityError(RuntimeError):
+  pass
+
+
+def _emulator():
+  try:
+    if QCOM:
+      from test.mockgpu.qcom import emu
+    else:
+      from test.mockgpu.amd import emu
+  except ImportError as e:
+    raise ParityError(f"tinygrad's {'qcom' if QCOM else 'amd'} mock is not importable: put tinygrad_repo on PYTHONPATH") from e
+  if not hasattr(emu, "hw_unverified"):
+    raise ParityError(f"this tinygrad's emulator has no {ARCH} rounding model, its check frames would not match a {HARDWARE}")
+  return emu
+
+
+def unverified(header: dict, blob: memoryview) -> dict[str, int]:
+  emu, out = _emulator(), {}
+  for call in header["calls"]:
+    if call["kind"] != "program":
+      continue
+    if (arch := call["target"].get("arch", "")).split(",")[0] != ARCH:  # IMAGE builds append ",IMAGE_PITCH_ALIGNMENT=64"
+      out[f"arch={arch}"] = out.get(f"arch={arch}", 0) + 1
+      continue
+    for sig, n in emu.hw_unverified(bytes(blob[slice(*call["lib"])])).items():
+      out[sig] = out.get(sig, 0) + n
+  return out
+
+
+def _array(blob: memoryview, stored: dict) -> np.ndarray:
+  return np.frombuffer(blob[slice(*stored["data"])], dtype=np.dtype(stored["dtype"])).reshape(stored["shape"]).copy()
+
+
+def replay(header: dict, blob: memoryview, shared: dict) -> list[np.ndarray]:
+  from tinygrad import Tensor
+  check = header["check"]
+  jit = PlainJit(header, blob, None, verify=False, shared=shared)
+  kwargs: dict = {}
+  for path, stored, device in check["inputs"]:
+    tensor = Tensor(_array(blob, stored), device=_device(device, None)).realize()
+    kwargs[path[0]] = tensor if len(path) == 1 else {**kwargs.get(path[0], {}), path[1]: tensor}
+  result = jit(**kwargs)
+  outputs = kwargs[check["output_container"]].values() if check["output_container"] else result
+  return [t.numpy() for t in (outputs if isinstance(outputs, (list, tuple)) or check["output_container"] else [outputs])]
+
+
+def _parts_with_offsets(path: str) -> tuple[dict, dict]:
+  parts, _ = read_parts(path)
+  with open(path, "rb") as f:
+    head = f.read(len(MAGIC) + 8)
+    start = len(MAGIC) + 8 + struct.unpack_from("<Q", head, len(MAGIC))[0]
+    f.seek(len(MAGIC) + 8)
+    entries = json.loads(f.read(start - len(MAGIC) - 8))["parts"]
+  base = start + -start % ALIGNMENT
+  offsets = {tuple(e["key"]) if isinstance(e["key"], list) else e["key"]: base + e["blob"][0] for e in entries}
+  return parts, offsets
+
+
+def _gate(parts: dict) -> None:
+  missing: dict[str, int] = {}
+  for header, blob in parts.values():
+    if "check" not in header:
+      raise ParityError("artifact has no check frame")
+    for sig, n in unverified(header, blob).items():
+      missing[sig] = missing.get(sig, 0) + n
+  if missing:
+    raise ParityError(f"kernels use instructions the emulator is not verified against a {HARDWARE} for: " +
+                      ", ".join(f"{s} ({n}x)" for s, n in sorted(missing.items(), key=lambda kv: -kv[1])))
+
+
+def check(path: str) -> None:
+  parts, _ = read_parts(path)
+  if ON_MOCK:
+    _gate(parts)
+  shared: dict = {}
+  for key, (header, blob) in parts.items():
+    got = replay(header, blob, shared)
+    for i, (g, stored) in enumerate(zip(got, header["check"]["outputs"], strict=True)):
+      want = _array(blob, stored)
+      if g.shape != want.shape or not np.array_equal(g.view(np.uint8), want.view(np.uint8)):
+        n = int(np.count_nonzero(g != want)) if g.shape == want.shape else -1
+        raise ParityError(f"part {key} output {i}: the {'mock' if ON_MOCK else 'device'}'s check frame differs from the stored one in {n} of {want.size} values")
+
+
+def record(path: str, out: str) -> None:
+  parts, offsets = _parts_with_offsets(path)
+  _gate(parts)
+  shared, writes = {}, []
+  for key, (header, blob) in parts.items():
+    for g, stored in zip(replay(header, blob, shared), header["check"]["outputs"], strict=True):
+      want = np.dtype(stored["dtype"])
+      if tuple(g.shape) != tuple(stored["shape"]) or g.dtype != want:
+        raise ParityError(f"part {key}: replayed output {g.dtype}{g.shape} does not fit the stored {want}{tuple(stored['shape'])}")
+      writes.append((offsets[key] + stored["data"][0], np.ascontiguousarray(g).tobytes()))
+  shutil.copyfile(path, out)
+  with open(out, "r+b") as f:
+    for offset, data in writes:
+      f.seek(offset)
+      f.write(data)
+
+
+def main(argv: list[str]) -> int:
+  if len(argv) < 2 or argv[0] not in ("check", "record") or (argv[0] == "record" and len(argv) != 3):
+    print("usage: mock_parity check ARTIFACT... | record ARTIFACT OUT", file=sys.stderr)
+    return 2
+  try:
+    if argv[0] == "record":
+      st = time.perf_counter()
+      record(argv[1], argv[2])
+      print(f"{argv[2]}: check frame recorded on the mock in {time.perf_counter() - st:.0f}s")
+      return 0
+    for path in argv[1:]:
+      st = time.perf_counter()
+      check(path)
+      print(f"{os.path.basename(path)}: bit-identical to its {HARDWARE} check frame on the {'mock' if ON_MOCK else 'device'} ({time.perf_counter() - st:.0f}s)")
+    return 0
+  except ParityError as e:
+    print(f"PARITY FAIL: {e}", file=sys.stderr)
+    return 1
+
+
+if __name__ == "__main__":
+  sys.exit(main(sys.argv[1:]))

openpilot/sunnypilot/modeld_v2/plain_artifact_compiler.py

@@ -80,7 +80,8 @@ def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict,
     if ast.op is Ops.PROGRAM:
       elf, info = ast.to_elf(), ast.arg
       global_size, local_size = info.launch_dims({})
-      calls.append({"kind": "program", "name": elf.name, "lib": blob.add(bytes(elf.lib)), "target": dataclasses.asdict(elf.target),
+      device_target = dataclasses.asdict(dataclasses.replace(elf.target, interface=elf.target.interface.removeprefix("MOCK")))
+      calls.append({"kind": "program", "name": elf.name, "lib": blob.add(bytes(elf.lib)), "target": device_target,
                     "signature": kernel_signature(ast, elf),
                     "global_size": [int(x) for x in global_size], "local_size": [int(x) for x in local_size],
                     "globals": list(info.globals), "outs": list(info.outs), "ins": list(info.ins), "args": arguments})

openpilot/sunnypilot/modeld_v2/tests/test_convert_upstream_pkl.py

@@ -6,6 +6,7 @@ See the LICENSE.md file in the root directory for more details.
 """
 import os
 import re
+import struct
 import sys
 import tempfile
 import unittest
@@ -25,8 +26,8 @@ if sys.platform == "darwin" and os.environ.get("DEV", "").startswith("MOCKUSB"):
 
 from openpilot.common.test import OpenpilotTestCase
 from openpilot.sunnypilot.modeld_v2 import convert_upstream_pkl
-from openpilot.sunnypilot.modeld_v2.convert_upstream_pkl import convert
-from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, load_model
+from openpilot.sunnypilot.modeld_v2.convert_upstream_pkl import _elf, convert
+from openpilot.sunnypilot.modeld_v2.helpers import ALIGNMENT, PlainJit, load_model
 from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame
 
 IMG, DESIRE, STATE, OUT = (6, 8, 8), (8,), (4, 16), (1, 32)
@@ -119,3 +120,18 @@ class TestConvertUpstreamPkl(OpenpilotTestCase):
     bundle = load_model(out, cache=False, verify=True)
     assert bundle["metadata"] == {"name": "tiny"} and bundle["input_specs"] == self.input_specs and bundle["output_specs"] == self.output_specs
     self.assert_matches_eager(bundle["run"], 6)
+
+
+class TestKernelElf(OpenpilotTestCase):
+  def test_sections_follow_the_code(self):
+    from tinygrad.runtime.support.elf import elf_loader
+    code = bytes(range(1, 129))
+    for rodata, entry in ((len(code), 0), (0, 256)):
+      image = bytearray(code + bytes(64) if entry < rodata else bytes(256) + code)
+      struct.pack_into("<q", image, rodata + 16, entry - rodata)
+      loaded, sections, _ = elf_loader(_elf(bytes(image), rodata, entry))
+      headers = {section.name: section.header for section in sections}
+      assert bytes(loaded[ALIGNMENT:]) == bytes(image), "the uploaded program differs from the pickle's"
+      assert headers[".rodata"].sh_addr == ALIGNMENT + rodata, "the loader reads the kernel descriptor at .rodata"
+      text = headers[".text"]
+      assert text.sh_addr <= ALIGNMENT + entry < text.sh_addr + text.sh_size, "the amd mock sizes the code from the entry to the end of .text"

release/ci/model_generator.py

@@ -13,26 +13,41 @@ from pathlib import Path
 from datetime import datetime, UTC
 
 
+_VERSION_REGEX = re.compile(r'^(?:\d+[a-zA-Z]+|\d+[vVbB]\d+|[vVbB]\d+|\d{4})$')
+_DECIMAL_VERSION_REGEX = re.compile(r'^([vVbB]?)(\d+(?:\.\d+)+)$')
+
+
+def _version(word: str) -> str | None:
+  match = _DECIMAL_VERSION_REGEX.match(word.strip('()[]{}.,;:'))
+  if match:
+    return match.group(1).upper() + match.group(2)
+  stripped = re.sub(r'[^a-zA-Z0-9]', '', word)
+  return stripped.upper() if _VERSION_REGEX.match(stripped) else None
+
+
+def _truncate(short_name: str) -> str:
+  return short_name[:8].rstrip('.')
+
+
 def create_short_name(full_name: str) -> str:
-  # Remove parentheses and extract alphanumeric words
+  # Remove parentheses and extract alphanumeric words, keeping decimals for versioning
   clean_name = re.sub(r'\([^)]*\)', '', full_name)
-  words = [re.sub(r'[^a-zA-Z0-9]', '', word) for word in clean_name.split() if re.sub(r'[^a-zA-Z0-9]', '', word)]
+  words = [re.sub(r'[^a-zA-Z0-9.]', '', word) for word in clean_name.split()]
+  words = [word for word in words if word]
 
   if len(words) == 1:
-    return words[0][:8].upper()
+    return _truncate((_version(words[0]) or words[0]).upper())
 
-  # Handle special case: Name + Version (e.g., "Word A1" -> "WordA1")
-  if len(words) == 2 and re.match(r'^[A-Za-z]\d+$', words[1]):
-    return (words[0] + words[1])[:8].upper()
+  # Handle special case: Name + Version (e.g., "Word A1" -> "WordA1", "Model v1.1" -> "ModelV1.1")
+  if len(words) == 2 and re.match(r'^[A-Za-z]?\d+$', words[1]):
+    return _truncate((words[0] + (_version(words[1]) or words[1])).upper())
 
   result = ""
   for word in words:
     # Version or number patterns
-    if (re.match(r'^\d+[a-zA-Z]+$', word) or
-        re.match(r'^\d+[vVbB]\d+$', word) or
-        re.match(r'^[vVbB]\d+$', word) or
-        re.match(r'^\d{4}$', word)):
-      result += word.upper()
+    version = _version(word)
+    if version:
+      result += version
     # All uppercase abbreviations (2-3 letters)
     elif re.match(r'^[A-Z]{2,3}$', word):
       result += word
@@ -45,7 +60,7 @@ def create_short_name(full_name: str) -> str:
       result += word
     else:
       result += word[0].upper()
-  return result[:8]
+  return _truncate(result)
 
 
 def create_pkl_name(full_name: str) -> str: