fix ci

these models will fail, look at workflow, likely need env setup now that we use recursion

.github/workflows/build-default-models.yaml

@@ -117,20 +117,6 @@ jobs:
             ${{ github.workspace }}/${{ env.SMALL_PKL }} \
             --device-input "*" --out-of-band --benchmark-runs 1
 
-      - name: Chunk small pkl
-        run: |
-          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
-          export PYTHONPATH=${{ github.workspace }}
-          python3 -c "
-          from openpilot.common.file_chunker import chunk_file, get_chunk_targets
-          import os
-          pkl = '${{ github.workspace }}/${{ env.SMALL_PKL }}'
-          size = os.path.getsize(pkl)
-          targets = get_chunk_targets(pkl, size)
-          chunk_file(pkl, targets)
-          print(f'Chunked into {len(targets)} files')
-          "
-
       - name: Prepare output
         env:
           MODEL_NAME: ${{ needs.resolve.outputs.safe_model_name }}
@@ -142,8 +128,7 @@ jobs:
           PKL_BASE="driving_tinygrad.pkl"
           mkdir -p "$OUTPUT_DIR"
 
-          cp "$MODELS_DIR/${PKL_BASE}".chunk* "$OUTPUT_DIR/"
-          cp "$MODELS_DIR/${PKL_BASE}.chunkmanifest" "$OUTPUT_DIR/"
+          cp "$MODELS_DIR/${PKL_BASE}" "$OUTPUT_DIR/"
 
           ONNX_SHA256=$(git cat-file blob "$(git ls-files -s "${{ github.workspace }}/${{ env.SMALL_ONNX }}" | awk '{print $2}')" | grep '^oid sha256:' | cut -d: -f2)
 
@@ -222,20 +207,6 @@ jobs:
               raise RuntimeError('Chestnut PCIe link not ready after 10 attempts')
           "
 
-      - name: Chunk big pkl
-        run: |
-          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
-          export PYTHONPATH=${{ github.workspace }}
-          python3 -c "
-          from openpilot.common.file_chunker import chunk_file, get_chunk_targets
-          import os
-          pkl = '${{ github.workspace }}/${{ env.BIG_PKL }}'
-          size = os.path.getsize(pkl)
-          targets = get_chunk_targets(pkl, size)
-          chunk_file(pkl, targets)
-          print(f'Chunked into {len(targets)} files')
-          "
-
       - name: Compile big warp
         run: |
           source ${UV_PROJECT_ENVIRONMENT}/bin/activate
@@ -275,8 +246,7 @@ jobs:
           PKL_BASE="big_driving_tinygrad.pkl"
           mkdir -p "$OUTPUT_DIR"
 
-          cp "$MODELS_DIR/${PKL_BASE}".chunk* "$OUTPUT_DIR/"
-          cp "$MODELS_DIR/${PKL_BASE}.chunkmanifest" "$OUTPUT_DIR/"
+          cp "$MODELS_DIR/${PKL_BASE}" "$OUTPUT_DIR/"
 
           PKL_LFS_SHA256=$(git cat-file blob "$(git ls-files -s "${{ github.workspace }}/${{ env.BIG_PKL }}" | awk '{print $2}')" | grep '^oid sha256:' | cut -d: -f2)
 
@@ -391,24 +361,14 @@ jobs:
           from datetime import datetime, UTC
 
           dm_dir = Path('dm_output')
-          manifest = list(dm_dir.glob('*.chunkmanifest'))
-          assert manifest, 'No chunkmanifest found'
-          pkl_name = manifest[0].name.removesuffix('.chunkmanifest')
-          num_chunks = int(manifest[0].read_text().strip())
-
-          chunks = []
-          for i in range(num_chunks):
-              chunk = dm_dir / f'{pkl_name}.chunk{i+1:02d}of{num_chunks:02d}'
-              chunks.append({
-                  'file_name': chunk.name,
-                  'sha256': hashlib.sha256(chunk.read_bytes()).hexdigest()
-              })
+          matches = sorted(dm_dir.glob('dmonitoring_*_tinygrad.pkl'))
+          assert len(matches) == 1, f'expected one DM pkl in {dm_dir}, found {matches}'
+          pkl_name = matches[0].name
 
           digest = hashlib.sha256()
-          for c in chunks:
-              with open(dm_dir / c['file_name'], 'rb') as f:
-                  while block := f.read(1024*1024):
-                      digest.update(block)
+          with open(dm_dir / pkl_name, 'rb') as f:
+              while block := f.read(1024*1024):
+                  digest.update(block)
 
           metadata = {
               'bundles': [{
@@ -422,8 +382,7 @@ jobs:
                       'type': 'chunked',
                       'artifact': {
                           'file_name': pkl_name,
-                          'download_uri': {'url': '', 'sha256': digest.hexdigest()},
-                          'chunks': chunks
+                          'download_uri': {'url': '', 'sha256': digest.hexdigest()}
                       }
                   }]
               }]
@@ -486,20 +445,6 @@ jobs:
             ${{ github.workspace }}/${{ env.DM_ONNX }} \
             ${{ github.workspace }}/${{ env.DM_PKL }}
 
-      - name: Chunk DM pkl
-        run: |
-          source ${UV_PROJECT_ENVIRONMENT}/bin/activate
-          export PYTHONPATH=${{ github.workspace }}
-          python3 -c "
-          from openpilot.common.file_chunker import chunk_file, get_chunk_targets
-          import os
-          pkl = '${{ github.workspace }}/${{ env.DM_PKL }}'
-          size = os.path.getsize(pkl)
-          targets = get_chunk_targets(pkl, size)
-          chunk_file(pkl, targets)
-          print(f'Chunked {pkl} into {len(targets)} chunks')
-          "
-
       - name: Compile DM warp
         run: |
           source ${UV_PROJECT_ENVIRONMENT}/bin/activate
@@ -532,8 +477,7 @@ jobs:
       - name: Prepare DM output
         run: |
           mkdir -p dm_output
-          cp ${{ github.workspace }}/${{ env.DM_PKL }}.chunk* dm_output/
-          cp ${{ github.workspace }}/${{ env.DM_PKL }}.chunkmanifest dm_output/
+          cp ${{ github.workspace }}/${{ env.DM_PKL }} dm_output/
           cp ${{ github.workspace }}/openpilot/selfdrive/modeld/models/dm_warp_* dm_output/
 
       - name: Upload DM artifact

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

@@ -137,6 +137,7 @@ jobs:
     needs: build_model
     runs-on: ubuntu-latest
     permissions:
+      actions: read
       id-token: write
       contents: write
     steps:
@@ -179,18 +180,21 @@ jobs:
           echo "artifact_name=$ARTIFACT_NAME" >> $GITHUB_OUTPUT
 
       - name: Download and extract model artifact
-        uses: actions/download-artifact@v4
-        with:
-          name: ${{ steps.read-artifact-name.outputs.artifact_name }}
-          path: output
+        env:
+          GH_TOKEN: ${{ github.token }}
+          ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
+        run: |
+          id=$(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: Create models folder
         env:
           ARTIFACT_NAME: ${{ steps.read-artifact-name.outputs.artifact_name }}
         run: |
           mkdir -p "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}"
-          cp -r output/* "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/"
-          rm -f "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/artifact_name.txt"
+          cp output/metadata.json "local_models/${RECOMPILED_DIR}/${ARTIFACT_NAME}/"
 
       - name: Upload to Hugging Face
         env:

.github/workflows/compile-warp-kernels.yaml

@@ -2,8 +2,6 @@ name: Compile warp for sunnypilot modeld
 
 on:
   workflow_dispatch:
-  schedule:
-    - cron: '0 0 * * 0'
 
 jobs:
   compile_warps:
@@ -43,13 +41,13 @@ jobs:
 
             WARP_PKL="openpilot/sunnypilot/modeld_v2/models/driving_warp_${res}_tinygrad.pkl"
             echo "Compiling $WARP_PKL on QCOM"
-            taskset -c 7 env DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1 python3 tinygrad_repo/examples/openpilot/compile_warp.py \
-              --frame ${FRAME_PARAM} --warp-to ${MODEL_SIZE} --layout yuv420 --frames 2 --output "${WARP_PKL}"
+            taskset -c 7 env DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1 python3 openpilot/sunnypilot/modeld_v2/compile_warp.py \
+              --frame ${FRAME_PARAM} --warp-to ${MODEL_SIZE} --output "${WARP_PKL}"
 
             BIG_WARP_PKL="openpilot/sunnypilot/modeld_v2/models/big_driving_warp_${res}_tinygrad.pkl"
             echo "Compiling $BIG_WARP_PKL on AMD"
-            taskset -c 7 env DEBUG=1 DEV=USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32 python3 tinygrad_repo/examples/openpilot/compile_warp.py \
-              --frame ${FRAME_PARAM} --warp-to ${MODEL_SIZE} --layout yuv420 --frames 2 --output "${BIG_WARP_PKL}"
+            taskset -c 7 env DEV=USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32 python3 openpilot/sunnypilot/modeld_v2/compile_warp.py \
+              --frame ${FRAME_PARAM} --warp-to ${MODEL_SIZE} --output "${BIG_WARP_PKL}"
           done
 
       - name: Re-enable powersave

.github/workflows/download-hf-model/action.yml

@@ -1,5 +1,5 @@
-name: Download HF model chunks
-description: Resolve and download model chunks from HuggingFace in parallel
+name: Download HF model
+description: Resolve and download models from HuggingFace in parallel
 
 inputs:
   hf_repo:
@@ -19,7 +19,7 @@ inputs:
 runs:
   using: composite
   steps:
-    - name: Download model chunks
+    - name: Download models
       shell: bash
       env:
         HF_REPO: ${{ inputs.hf_repo }}
@@ -33,7 +33,7 @@ runs:
         resolve_chunks() {
           local HF_PATH="$1" ONNX_HASH="$2" CANONICAL="$3" DEST_DIR="$4"
           local JSON_URL="https://huggingface.co/datasets/${HF_REPO}/resolve/main/${HF_PATH}/default_models.json"
-          local DEFAULTS BUNDLE ARTIFACT BASE_URL NUM_CHUNKS
+          local DEFAULTS BUNDLE ARTIFACT FILE_URL BASE_URL
           DEFAULTS=$(curl -fsSL "$JSON_URL")
           if [ -n "$TINYGRAD_REF" ]; then
             BUNDLE=$(echo "$DEFAULTS" | jq --arg hash "$ONNX_HASH" --arg tg "$TINYGRAD_REF" '
@@ -51,20 +51,12 @@ runs:
           fi
 
           ARTIFACT=$(echo "$BUNDLE" | jq -r '.models[0].artifact')
-          BASE_URL=$(echo "$ARTIFACT" | jq -r '.download_uri.url' | sed 's|/[^/]*$||')
-          NUM_CHUNKS=$(echo "$ARTIFACT" | jq -r '.chunks | length')
+          FILE_URL=$(echo "$ARTIFACT" | jq -r '.download_uri.url')
+          BASE_URL=$(echo "$FILE_URL" | sed 's|/[^/]*$||')
 
           mkdir -p "$DEST_DIR"
-          while IFS= read -r CHUNK_NAME; do
-            CHUNK_IDX=$(echo "$CHUNK_NAME" | grep -oP 'chunk\K[0-9]+of[0-9]+' || true)
-            if [ -z "$CHUNK_IDX" ]; then
-              echo "::error::Failed to parse chunk index from: $CHUNK_NAME"
-              return 1
-            fi
-            ENCODED_URL=$(python3 -c "import urllib.parse; print(urllib.parse.quote('${BASE_URL}/${CHUNK_NAME}', safe=':/'))")
-            printf '%s\t%s\n' "$ENCODED_URL" "${DEST_DIR}/${CANONICAL}.chunk${CHUNK_IDX}" >> "$DOWNLOAD_LIST"
-          done < <(echo "$ARTIFACT" | jq -r '.chunks[].file_name')
-          echo "$NUM_CHUNKS" > "${DEST_DIR}/${CANONICAL}.chunkmanifest"
+          ENCODED_URL=$(python3 -c "import urllib.parse; print(urllib.parse.quote('${FILE_URL}', safe=':/'))")
+          printf '%s\t%s\n' "$ENCODED_URL" "${DEST_DIR}/${CANONICAL}" >> "$DOWNLOAD_LIST"
 
           if [ "$CANONICAL" = "dmonitoring_model_tinygrad.pkl" ]; then
             for warp in dm_warp_1928x1208_tinygrad.pkl dm_warp_1344x760_tinygrad.pkl; do
@@ -87,7 +79,7 @@ runs:
         done
 
         TOTAL=$(wc -l < "$DOWNLOAD_LIST")
-        echo "Downloading $TOTAL chunks with 8 parallel connections..."
+        echo "Downloading $TOTAL files with 8 parallel connections..."
         xargs -P8 -d'\n' -I{} bash -c '
           URL="${1%%	*}"
           DEST="${1#*	}"

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

@@ -102,56 +102,39 @@ jobs:
           cat $GITHUB_OUTPUT
 
       - name: Fetch Models
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
         run: |
           cd ${{ github.workspace }}/openpilot/openpilot
+          fetch_lfs() {
+            sudo apt-get install -y -qq aria2 > /dev/null # Parallel ranges because gh is flaky
+            LFS_URL=$(git config -f "$(git rev-parse --show-toplevel)/.lfsconfig" lfs.url)
+            for f in "$@"; do
+              grep -qs "^version https://git-lfs" "$f" || continue
+              OID=$(awk -F'sha256:' '/^oid/{print $2}' "$f"); SIZE=$(awk '/^size/{print $2}' "$f")
+              HREF=$(curl -fsSL --retry 3 -X POST -H "Accept: application/vnd.git-lfs+json" -H "Content-Type: application/vnd.git-lfs+json" "$LFS_URL/objects/batch" \
+                -d "{\"operation\":\"download\",\"transfer\":[\"basic\"],\"objects\":[{\"oid\":\"$OID\",\"size\":$SIZE}]}" | jq -r '.objects[0].actions.download.href')
+              aria2c -x 16 -s 16 -k 16M --max-tries=5 --retry-wait=5 --file-allocation=none --summary-interval=30 --allow-overwrite=true \
+                -d "$(dirname "$f")" -o "$(basename "$f")" "$HREF"
+              echo "$OID  $f" | sha256sum -c -
+            done
+          }
+          UPSTREAM_PKL=selfdrive/modeld/models/big_driving_tinygrad.pkl
           if [ "${{ inputs.target_hardware }}" != "chestnut" ]; then
-            ONNX_COUNT=$(find selfdrive/modeld/models -maxdepth 1 -type f -name "*.onnx" ! -name "big_*.onnx" ! -name "dmonitoring_*.onnx" 2>/dev/null | wc -l)
+            git lfs pull -X "**/selfdrive/modeld/models/big_*.onnx,**/selfdrive/modeld/models/dmonitoring_*.onnx,**/selfdrive/modeld/models/big_*.pkl,**/selfdrive/modeld/models/dmonitoring_*.pkl"
+            rm -f selfdrive/modeld/models/big_*.onnx selfdrive/modeld/models/dmonitoring_*.onnx selfdrive/modeld/models/big_*.pkl selfdrive/modeld/models/dmonitoring_*.pkl
+          elif compgen -G "selfdrive/modeld/models/big_*.onnx" > /dev/null; then
+            fetch_lfs selfdrive/modeld/models/big_*.onnx selfdrive/modeld/models/big_*.pkl
+            find selfdrive/modeld/models -type f \( -name "*.onnx" -o -name "*.pkl" \) ! -name "big_*.onnx" ! -name "big_*.pkl" -delete
+          elif [ -f "$UPSTREAM_PKL" ]; then
+            echo "No big model ONNX tracked: converting commaai's precompiled $UPSTREAM_PKL, with the tinygrad it was built with"
+            fetch_lfs "$UPSTREAM_PKL"
+            find selfdrive/modeld/models -type f \( -name "*.onnx" -o -name "*.pkl" \) ! -path "$UPSTREAM_PKL" -delete
+            GITMODULES="$(git rev-parse --show-toplevel)/.gitmodules"
+            TG_NAME=$(git config -f "$GITMODULES" --get-regexp '^submodule\..*\.path$' | awk '$2 == "tinygrad_repo" {sub(/^submodule\./, "", $1); sub(/\.path$/, "", $1); print $1}')
+            TG_URL=$(git config -f "$GITMODULES" "submodule.$TG_NAME.url" | sed 's|^\.\./\.\./|https://github.com/|')
+            echo "$TG_URL $(git rev-parse HEAD:tinygrad_repo) $(sha256sum "$UPSTREAM_PKL" | cut -d' ' -f1)" > selfdrive/modeld/models/upstream_pkl.txt
           else
-            ONNX_COUNT=$(find selfdrive/modeld/models -maxdepth 1 -type f -name "big_*.onnx" 2>/dev/null | wc -l)
-          fi
-          if [ "$ONNX_COUNT" -eq 0 ]; then
-            echo "Detected NEW stock architecture (no ONNX files tracked). Fetching from HF and skipping LFS pull."
-            COMMIT_MSG=$(git log -1 --pretty=%B)
-            HF_HASH=$(echo "$COMMIT_MSG" | grep -oE '[a-f0-9]{8}' | head -n 1)
-
-            if [ -n "$HF_HASH" ]; then
-              echo "Downloading ONNX for hash $HF_HASH from HuggingFace."
-              pip install huggingface_hub
-              rm -f selfdrive/modeld/models/*.onnx
-              python3 - <<EOF
-          import sys, os, shutil
-          from huggingface_hub import HfApi, hf_hub_download
-          short_hash = "$HF_HASH"
-          target_dir = "selfdrive/modeld/models"
-          repo = "commaai/openpilot_driving_models"
-          api = HfApi()
-          files = list(api.list_repo_tree(repo, repo_type="model", recursive=True))
-          matching_folder = next((f.path.split("/")[0] for f in files if f.path.startswith(short_hash)), None)
-          if not matching_folder:
-            print(f"Could not find HF folder for hash {short_hash}")
-            sys.exit(1)
-          os.makedirs(target_dir, exist_ok=True)
-          for f in files:
-            if f.path.startswith(matching_folder) and f.path.endswith(".onnx"):
-              p = hf_hub_download(repo_id=repo, repo_type="model", filename=f.path)
-              shutil.copy2(p, os.path.join(target_dir, os.path.basename(f.path)))
-          EOF
-              rm -f selfdrive/modeld/models/*.pkl
-              touch selfdrive/modeld/models/new_arch.txt
-            else
-              echo "Could not parse HuggingFace hash from commit message."
-              exit 1
-            fi
-          else
-            if [ "${{ inputs.target_hardware }}" != "chestnut" ]; then
-              git lfs pull -X "**/selfdrive/modeld/models/big_*.onnx,**/selfdrive/modeld/models/dmonitoring_*.onnx,**/selfdrive/modeld/models/big_*.pkl,**/selfdrive/modeld/models/dmonitoring_*.pkl"
-              rm -f selfdrive/modeld/models/big_*.onnx selfdrive/modeld/models/dmonitoring_*.onnx selfdrive/modeld/models/big_*.pkl selfdrive/modeld/models/dmonitoring_*.pkl
-            else
-              git lfs pull -I "**/selfdrive/modeld/models/big_*.onnx,**/selfdrive/modeld/models/big_*.pkl" -X ""
-              find selfdrive/modeld/models -type f \( -name "*.onnx" -o -name "*.pkl" \) ! -name "big_*.onnx" ! -name "big_*.pkl" -delete
-            fi
+            echo "::error::no big model ONNX or precompiled big_driving_tinygrad.pkl at this ref"
+            exit 1
           fi
 
           if grep -lIF "version https://git-lfs.github.com/spec/v1" selfdrive/modeld/models/*.onnx selfdrive/modeld/models/*.pkl 2>/dev/null; then
@@ -162,10 +145,11 @@ jobs:
         uses: actions/upload-artifact@v4
         with:
           name: models-${{ env.REF }}${{ inputs.artifact_suffix  }}
+          compression-level: 0
           path: |
             ${{ github.workspace }}/openpilot/openpilot/selfdrive/modeld/models/*.onnx
             ${{ github.workspace }}/openpilot/openpilot/selfdrive/modeld/models/*.pkl
-            ${{ github.workspace }}/openpilot/openpilot/selfdrive/modeld/models/new_arch.txt
+            ${{ github.workspace }}/openpilot/openpilot/selfdrive/modeld/models/upstream_pkl.txt
           if-no-files-found: error
 
   build_model:
@@ -224,11 +208,11 @@ jobs:
           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 JIT_BATCH_SIZE=0 OPENPILOT_HACKS=1"
+          TG_FLAGS_QCOM="DEV=QCOM IMAGE=1 FLOAT16=1 NOLOCALS=1 OPENPILOT_HACKS=1"
           if [ "${{ inputs.target_hardware }}" == "chestnut" ]; then
             echo "CHESTNUT build"
             export CHESTNUT=1
-            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 JIT_BATCH_SIZE=0 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32"
+            TG_FLAGS="DEBUG=1 DEV=USB+AMD:LLVM FRAME_DEV=CPU FLOAT16=1 GMMU=0 TC_OPT=2 TC_MIN_GLOBALS=32"
             OUTPUT_PKL="${{ env.MODELS_DIR }}/big_driving_tinygrad.pkl"
           else
             echo "QCOM build"
@@ -238,20 +222,20 @@ jobs:
 
           NATIVE_PKL=$(find "${{ env.MODELS_DIR }}" -maxdepth 1 -name "*.pkl" -print -quit)
 
-          if [ -f "${{ env.MODELS_DIR }}/new_arch.txt" ]; then
-            echo "Detected new arch"
-            DOWNLOADED_ONNX=$(find "${{ env.MODELS_DIR }}" -maxdepth 1 -name "*driving*.onnx" -print -quit)
-            if [ -n "$DOWNLOADED_ONNX" ]; then
-              echo "Compiling ONNX: $DOWNLOADED_ONNX"
-              env ${TG_FLAGS} taskset -c 7 python3 openpilot/sunnypilot/modeld_v2/compile_modeld_v2.py \
-                "$DOWNLOADED_ONNX" "$OUTPUT_PKL" --device-input "*" --out-of-band --benchmark-runs 1
-
-              echo "Chunking pkl"
-              python3 -c "from openpilot.common.file_chunker import chunk_file, get_chunk_targets; import os; p='$OUTPUT_PKL'; chunk_file(p, get_chunk_targets(p, os.path.getsize(p)))"
-            else
-              echo "Failed to find downloaded ONNX."
-              exit 1
-            fi
+          if [ -f "${{ env.MODELS_DIR }}/upstream_pkl.txt" ]; then
+            read -r TG_URL TG_SHA _ < "${{ env.MODELS_DIR }}/upstream_pkl.txt"
+            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
+            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.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 ${TG_FLAGS} taskset -c 7 python3 openpilot/sunnypilot/modeld_v2/compile_modeld_v2.py \
+              "$NEW_ARCH_ONNX" "$OUTPUT_PKL" --device-input "*" --benchmark-runs 1
           else
             echo "Detected onnx architecture. Routing to native compile_modeld.py."
             # Detect model type and build compile args
@@ -317,8 +301,7 @@ jobs:
           # Copy the model files
           rsync -avm \
             --include='*.dlc' \
-            --include='*.chunk*' \
-            --include='*.chunkmanifest' \
+            --include='/${{ inputs.target_hardware == 'chestnut' && 'big_' || '' }}driving_tinygrad.pkl' \
             --exclude='*' \
             --delete-excluded \
             --chown=comma:comma \
@@ -329,6 +312,7 @@ jobs:
             --output-dir "${{ env.OUTPUT_DIR }}" \
             --custom-name "${{ env.MODEL_NAME }}" \
             --upstream-branch "${{ inputs.upstream_branch }}" \
+            $([ -f "${{ env.MODELS_DIR }}/upstream_pkl.txt" ] && echo "--onnx-sha256 $(cut -d' ' -f3 "${{ env.MODELS_DIR }}/upstream_pkl.txt")") \
             ${{ inputs.is_20hz && '--is-20hz' || '' }}
 
       - name: Write artifact name to file
@@ -340,6 +324,7 @@ jobs:
         with:
           name: model-${{ env.MODEL_NAME }}${{ inputs.artifact_suffix }}-${{ github.run_number }}
           path: ${{ env.OUTPUT_DIR }}
+          compression-level: 0
 
       - name: Upload artifact name file
         uses: actions/upload-artifact@v4

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

@@ -503,8 +503,8 @@ jobs:
           mkdir -p ${{ env.OUTPUT_DIR }}
           tar xzf prebuilt.tar.gz -C ${{ env.OUTPUT_DIR }}
 
-      - name: Download model chunks from HF
-        uses: ./.github/workflows/download-hf-model-chunks
+      - name: Download models from HF
+        uses: ./.github/workflows/download-hf-model
         with:
           hf_repo: sunnypilot/sunnypilot_models_v1
           dest_dir: ${{ env.OUTPUT_DIR }}/openpilot/selfdrive/modeld/models
@@ -573,8 +573,8 @@ jobs:
           mkdir -p ${{ env.OUTPUT_DIR }}
           tar xzf prebuilt.tar.gz -C ${{ env.OUTPUT_DIR }}
 
-      - name: Download model chunks from HF
-        uses: ./.github/workflows/download-hf-model-chunks
+      - name: Download models from HF
+        uses: ./.github/workflows/download-hf-model
         with:
           hf_repo: sunnypilot/sunnypilot_models_v1
           dest_dir: ${{ env.OUTPUT_DIR }}/openpilot/selfdrive/modeld/models

.github/workflows/test_models.yml

@@ -47,38 +47,22 @@ jobs:
         with:
           submodules: true
 
-      - name: Download Model Chunks in Parallel
+      - name: Download model
         env:
           HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          MODEL_URL: ${{ matrix.artifact.download_uri.url }}
+          MODEL_FILE: ${{ matrix.artifact.file_name }}
+        shell: bash
         run: |
-          mkdir -p /tmp/model_chunks
-          echo '${{ toJson(matrix.artifact.chunks) }}' > chunks.json
-
-          BASE_URL="${{ matrix.artifact.download_uri.url }}"
-          export BASE_DIR=$(dirname "$BASE_URL")
-
-          python3 -c '
-          import json, os
-          with open("chunks.json") as f:
-            chunks = json.load(f)
-          manifest_path = f"/tmp/model_chunks/${{ matrix.artifact.file_name }}.chunkmanifest"
-          with open(manifest_path, "w") as f:
-            f.write(str(len(chunks)))
-          base_dir = os.environ["BASE_DIR"]
-          hf_token = os.environ.get("HF_TOKEN", "")
-          with open("/tmp/curl_config.txt", "w") as f:
-            for c in chunks:
-              fn = c["file_name"]
-              f.write(f"url = \"{base_dir}/{fn}\"\noutput = \"/tmp/model_chunks/{fn}\"\n")
-              if hf_token:
-                f.write(f"header = \"Authorization: Bearer {hf_token}\"\n")
-          '
-          curl -Z --parallel-immediate --parallel-max 12 --retry 3 --retry-all-errors -s -S -f -L -K /tmp/curl_config.txt
+          mkdir -p /tmp/models
+          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 "/tmp/models/$MODEL_FILE" "$MODEL_URL"
 
       - name: Run Model Compatibility Test
         env:
           MODEL_BASE_NAME: ${{ matrix.artifact.file_name }}
-          MODEL_CHUNK_DIR: "/tmp/model_chunks"
+          MODEL_DIR: "/tmp/models"
         run: |
           export PYTHONPATH="$GITHUB_WORKSPACE/tinygrad_repo:$GITHUB_WORKSPACE:$PYTHONPATH"
           python3 -m pip uninstall -y tinygrad || uv pip uninstall -y tinygrad --system || true

openpilot/selfdrive/modeld/SConscript

@@ -55,7 +55,7 @@ def chestnut_action(command):
 def compile_model(onnx_path, pkl_path):
   onnx_path, target_pkl_path = File(onnx_path).abspath, File(pkl_path).abspath
   cmd = (f'{tg_flags} {mac_brew_string} {taskset}python3 "{compiler}/compile_onnx.py" '
-         f'"{onnx_path}" "{target_pkl_path}" --device-input "*" --out-of-band --benchmark-runs 1')
+         f'"{onnx_path}" "{target_pkl_path}" --out-of-band --benchmark-runs 1')
   lenv.Command(
     target_pkl_path,
     tinygrad_files + [onnx_path, Value(cmd)],

openpilot/selfdrive/modeld/dmonitoringmodeld.py

@@ -16,7 +16,6 @@ from openpilot.common.realtime import config_realtime_process
 from openpilot.common.transformations.model import dmonitoringmodel_intrinsics
 from openpilot.common.transformations.camera import _ar_ox_fisheye, _os_fisheye
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
-from openpilot.common.file_chunker import open_file_chunked
 from openpilot.selfdrive.modeld.parse_model_outputs import sigmoid, safe_exp
 
 SEND_RAW_PRED = os.getenv('SEND_RAW_PRED')
@@ -28,7 +27,7 @@ class ModelState:
   output: np.ndarray
 
   def __init__(self, cam_w: int, cam_h: int):
-    jits = load_oob(open_file_chunked(MODEL_PKL_PATH))
+    jits = load_oob(MODEL_PKL_PATH)
     self.DEV = jits['input_specs']['input_img'][2]
     self.input_shapes = jits['metadata']['input_shapes']
     self.output_slices = pickle.loads(base64.b64decode(jits['metadata']['metadata']['output_slices']))

openpilot/selfdrive/modeld/helpers.py

@@ -1,9 +1,12 @@
 import io
+import os
 import pickle
 import struct
+import sys
 from pathlib import Path
 
-from openpilot.common.file_chunker import get_manifest_path
+from openpilot.common.file_chunker import get_chunk_name, get_manifest_path, open_file_chunked
+from openpilot.common.hardware import AGNOS
 from openpilot.common.hardware.usb import CHESTNUT_USB_PRODUCT, USB_DEVICES_PATH, is_chestnut_usb_id
 
 MODELS_DIR = Path(__file__).resolve().parent / 'models'
@@ -13,15 +16,19 @@ def modeld_pkl_path(chestnut: bool):
   prefix = 'big_' if chestnut else ''
   return MODELS_DIR / f'{prefix}driving_tinygrad.pkl'
 
-def load_oob(f):
-  opcodes = f.read(struct.unpack('<q', f.read(8))[0])
-  def buffers():
-    while (h := f.read(8)):
-      pb = pickle.PickleBuffer(bytearray(struct.unpack('<q', h)[0]))
-      if f.readinto(pb) != pb.raw().nbytes:
-        raise EOFError("incomplete model buffer")
-      yield pb
-  return pickle.load(io.BytesIO(opcodes), buffers=buffers())
+def load_oob(path, chestnut=False):
+  from tinygrad import Context, Device
+  from tinygrad.device import Buffer
+  device = 'USB+AMD:LLVM' if chestnut else 'QCOM' if AGNOS else 'METAL' if sys.platform == 'darwin' else 'CPU:LLVM'
+  chunks = 0 if os.path.isfile(path) else int(Path(get_manifest_path(path)).read_text())
+  paths = [get_chunk_name(path, i, chunks) for i in range(chunks)] or [path]
+  with Context(DEV=device), open_file_chunked(path) as f:
+    opcodes = f.read(struct.unpack('<q', f.read(8))[0])
+    buffers, stage = Buffer(Device.DEFAULT, sum(map(os.path.getsize, paths)) - 8 - len(opcodes)).ensure_allocated(), bytearray(64 << 20)
+    for offset in range(0, buffers.nbytes, len(stage)):
+      n = f.readinto(piece := memoryview(stage)[:min(len(stage), buffers.nbytes - offset)])
+      buffers.view(n, offset).ensure_allocated().copy_from(Buffer("PYTHON", n, initial_value=piece[:n]))
+    return type("_Unpickler", (pickle.Unpickler,), {"persistent_load": lambda self, pid: buffers.view(*pid)})(io.BytesIO(opcodes)).load()
 
 def chestnut_present() -> bool:
   for d in USB_DEVICES_PATH.glob("*"):

openpilot/selfdrive/modeld/modeld.py

@@ -6,6 +6,7 @@ from functools import cached_property
 import os
 os.environ['GMMU'] = '0' # for chestnut fast loading, noop for qcom
 from tinygrad.device import Buffer, Device
+from tinygrad.engine.realize import lower_and_compile
 from tinygrad.dtype import DType, dtypes
 from tinygrad.tensor import Tensor
 from tinygrad.helpers import round_up
@@ -34,7 +35,6 @@ from openpilot.selfdrive.controls.lib.desire_helper import DesireHelper
 from openpilot.selfdrive.controls.lib.drive_helpers import get_accel_from_plan, should_stop, smooth_value, get_curvature_from_plan
 from openpilot.selfdrive.modeld.parse_model_outputs import Parser
 from openpilot.selfdrive.modeld.fill_model_msg import fill_model_msg, fill_driving_model_data, fill_pose_msg, PublishState
-from openpilot.common.file_chunker import open_file_chunked
 from openpilot.selfdrive.modeld.constants import ModelConstants, Plan
 from openpilot.selfdrive.modeld.helpers import MODELS_DIR, chestnut_present, chestnut_compiled, modeld_pkl_path, load_oob
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
@@ -134,8 +134,8 @@ class FrameMeta:
 
 
 def input_view(buffer: Buffer, shape: tuple[int, ...], dtype: DType, offset: int) -> Tensor:
-  view = buffer.view(math.prod(shape), dtype, offset).ensure_allocated()
-  return Tensor(UOp.from_buffer(view)).reshape(shape)
+  view = buffer.view(math.prod(shape) * dtype.itemsize, offset).ensure_allocated()
+  return Tensor(UOp.from_buffer(view, dtype)).reshape(shape)
 
 
 class ModelState(ModelStateBase):
@@ -143,7 +143,7 @@ class ModelState(ModelStateBase):
 
   def __init__(self, cam_w: int, cam_h: int, chestnut: bool):
     ModelStateBase.__init__(self)
-    jits = load_oob(open_file_chunked(modeld_pkl_path(chestnut)))
+    jits = load_oob(modeld_pkl_path(chestnut), chestnut)
     self.model_device = jits['input_specs']['new_img'][2]
     self.input_shapes = {name: (shape, np.dtype(dtype)) for name, (shape, dtype, _) in jits['input_specs'].items()}
     self.state_pairs = {name: f'next_{name}' for name in self.input_shapes if f'next_{name}' in jits['metadata']['output_shapes']}
@@ -158,7 +158,9 @@ class ModelState(ModelStateBase):
     self.pack_inputs()
     with open(MODELS_DIR / f'{"big_" if chestnut else ""}driving_warp_{cam_w}x{cam_h}_tinygrad.pkl', 'rb') as f:
       self.run_warp = pickle.load(f)['run']
+    self.run_warp.captured._linear = lower_and_compile(self.run_warp.captured._linear)
     self.run_model = jits['run']
+    self.run_model.captured._linear = lower_and_compile(self.run_model.captured._linear)
     self.outputs = {name: Tensor(np.zeros(shape, dtype=dtype), device=device).realize() for name, (shape, dtype, device) in jits['output_specs'].items()}
     for name, next_name in self.state_pairs.items():
       state = self.input_queues[name]
@@ -230,7 +232,8 @@ def main(demo=False):
 
   CHESTNUT = chestnut_present() and chestnut_compiled()
   if CHESTNUT:
-    os.environ['HCQDEV_WAIT_TIMEOUT_MS'] = '3000'
+    from tinygrad.runtime.ops_amd import AMDDevice
+    AMDDevice.wait_timeout_ms = 3000
   params = Params()
   params.put_bool("ChestnutLoading", CHESTNUT)
   params.remove("ChestnutActive")

openpilot/sunnypilot/modeld_v2/compile_modeld.py

@@ -7,14 +7,12 @@ See the LICENSE.md file in the root directory for more details.
 """
 
 import argparse
-import atexit
 import codecs
 import gc
 import math
 import multiprocessing as mp
 import os
 import pickle
-import shutil
 import tempfile
 import time
 from functools import partial
@@ -22,8 +20,8 @@ from typing import Any
 
 import numpy as np
 
-from openpilot.common.file_chunker import chunk_file, get_chunk_targets, open_file_chunked
-from openpilot.sunnypilot.modeld_v2.helpers import dump_oob, load_oob
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, patch_fetch_fw, read_parts, write_bundle
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame, record_kernel_linear, serialize_jit
 import openpilot.sunnypilot.modeld_v2.stock_dependencies as stock
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 
@@ -36,34 +34,7 @@ from tinygrad.tensor import Tensor
 
 os.environ['GMMU'] = '0'
 
-def _patch_tinygrad_fetch_fw():
-  try:
-    import hashlib
-    import pathlib
-    import zstandard
-    from tinygrad import helpers
-    _orig_fetch_fw = helpers.fetch_fw
-    def fetch_fw(path, name, sha256):
-      p = pathlib.Path(f"/lib/firmware/{path}/{name}.zst")
-      if p.is_file():
-        blob = zstandard.ZstdDecompressor().stream_reader(p.read_bytes()).read()
-        if hashlib.sha256(blob).hexdigest() == sha256:
-          return blob
-      return _orig_fetch_fw(path, name, sha256)
-    helpers.fetch_fw = fetch_fw
-  except ImportError:
-    pass
-_patch_tinygrad_fetch_fw()
-
-def _patch_ops_serialization():
-  try:
-    from tinygrad.uop.ops import Ops
-    def __reduce_ex__(self, proto):
-      return (getattr, (self.__class__, self.name))
-    Ops.__reduce_ex__ = __reduce_ex__
-  except ImportError:
-    pass
-_patch_ops_serialization()
+patch_fetch_fw()
 
 MODEL_TYPES = ('vision_policy', 'supercombo', 'vision_multi_policy')
 WARP_INPUTS = ['tfm', 'big_tfm']
@@ -72,6 +43,7 @@ POLICY_INPUTS = ['img_q', 'big_img_q', 'feat_q', 'desire_q', 'packed_npy_inputs'
 class MetadataOnnxPBParser(OnnxPBParser):
   def _parse_ModelProto(self) -> dict:
     obj: dict[str, Any] = {"graph": {"input": [], "output": []}, "metadata_props": []}
+    self.opset_imports = {}
     for fid, wire_type in self._parse_message(self.reader.len):
       match fid:
         case 7:
@@ -150,6 +122,10 @@ def get_policy_npy_shapes(input_shapes: dict, is_supercombo: bool = False) -> tu
   return shapes, sizes
 
 
+def host_view(tensor: Tensor) -> np.ndarray: # QCOM -> CPU round trip of tfm takes 5ms longer than host memory view
+  return np.frombuffer(tensor._buffer()._host_mv(), dtype=np.float32).reshape(tensor.shape)
+
+
 def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = Device.DEFAULT,
                             is_supercombo: bool = False) -> tuple[dict, dict]:
   road_key, _ = _detect_vision_keys(input_shapes)
@@ -185,7 +161,7 @@ def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = D
     '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=device if os.getenv('CHESTNUT') else 'NPY').realize(),
+    'packed_npy_inputs': Tensor(packed_npy_inputs, device=device).realize(),
   }
 
   if features_buffer:
@@ -197,9 +173,11 @@ def generate_queues_and_npy(input_shapes: dict, frame_skip: int, device: str = D
   for key in ('tfm', 'big_tfm'):
     if key not in npy_arrays:
       npy_arrays[key] = np.eye(3, dtype=np.float32)
-    queues[key] = Tensor(npy_arrays[key], device='NPY').realize()
+    queues[key] = Tensor(npy_arrays[key], device=device).realize()
 
-  return queues, npy_arrays
+  packed = host_view(queues['packed_npy_inputs'])
+  npy_arrays |= {k: v.reshape(s) for (k, s), v in zip(shapes.items(), np.split(packed, split_indices) if sizes else [], strict=True)}
+  return queues, npy_arrays | {k: host_view(queues[k]) for k in ('tfm', 'big_tfm')}
 
 
 def make_split_input_queues(vision_input_shapes: dict, policy_input_shapes: dict,
@@ -221,8 +199,8 @@ def make_warp_queues(device=Device.DEFAULT):
     '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
+  queues = {k: Tensor(v, device=device).realize() for k, v in npy.items()}
+  return queues, {k: host_view(v) for k, v in queues.items()}
 
 
 def make_run_policy(vision_runner, policy_runners: list, features_slice: slice, frame_skip: int, input_shapes: dict):
@@ -282,8 +260,19 @@ def make_run_policy(vision_runner, policy_runners: list, features_slice: slice,
   return run_policy
 
 
+class _CheckFrameRecorder:
+  def __init__(self, jit):
+    self.jit, self.frame = jit, None
+
+  def __call__(self, *args, **kwargs):
+    if self.frame is not None or args or self.jit.cnt < 2:
+      return self.jit(*args, **kwargs)
+    self.frame, result = capture_frame(self.jit, kwargs, None)
+    return result
+
+
 def compile_jit(jit_or_fn, input_keys_or_make_inputs, make_queues=None, make_random_inputs=None,
-                benchmark_runs: int = 1, clear_refs=None, return_loaded: bool = True):
+                benchmark_runs: int = 1, clear_refs=None):
   if callable(input_keys_or_make_inputs) and make_queues is None:
     fn = jit_or_fn
     make_inputs = input_keys_or_make_inputs
@@ -334,22 +323,23 @@ def compile_jit(jit_or_fn, input_keys_or_make_inputs, make_queues=None, make_ran
 
   print('capture + replay')
   gc.collect()
-  test_val = run_eval(jit, 42, 3)
-  print(f'pickle round trip ({benchmark_runs} runs per seed)')
-  with tempfile.TemporaryFile(dir=".") as f:
-    dump_oob(jit, f)
-    if clear_refs:
-      clear_refs()
-    gc.collect()
-    f.seek(0)
-    loaded_jit = load_oob(f)
+  recorder = _CheckFrameRecorder(jit)
+  with record_kernel_linear() as recorded:
+    test_val = run_eval(recorder, 42, 3)
+  assert len(recorded) == 1, f"expected one captured kernel list, got {len(recorded)}"
+  part = serialize_jit(jit, recorded[0], recorder.frame)
+  print(f'plain artifact round trip ({benchmark_runs} runs per seed)')
+  if clear_refs:
+    clear_refs()
+  gc.collect()
+  loaded_jit = PlainJit(part[0], memoryview(part[1]))
 
   for seed in (42, 43):
     reference = test_val if seed == 42 else run_eval(jit_or_fn, seed, 1)
     actual = run_eval(loaded_jit, seed, benchmark_runs)
     for ref, val in zip(reference, actual, strict=True):
       np.testing.assert_array_equal(ref, val)
-  return loaded_jit if return_loaded else jit
+  return part
 
 
 def _parse_size(size_str: str) -> tuple[int, int]:
@@ -357,16 +347,6 @@ def _parse_size(size_str: str) -> tuple[int, int]:
   return int(width), int(height)
 
 
-def read_file_chunked_to_disk(path):
-  if not path:
-    return None
-  tmp_path = f'{path}.unchunked'
-  with open(tmp_path, 'wb') as f, open_file_chunked(path) as src:
-    shutil.copyfileobj(src, f)
-  atexit.register(lambda: os.path.exists(tmp_path) and os.remove(tmp_path))
-  return tmp_path
-
-
 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)]:
@@ -387,10 +367,7 @@ def _compile_warp_resolution_worker(cam_w, cam_h, model_w, model_h, benchmark_ru
     warp = None
   compiled_jit = compile_jit(warp, WARP_INPUTS, make_warp_queues, make_random_inputs=make_random_warp_inputs,
                             benchmark_runs=benchmark_runs, clear_refs=cleanup_warp)
-
-  result_data = (compiled_jit, Device.DEFAULT)
-  with open(result_path, "wb") as f:
-    dump_oob(result_data, f)
+  write_bundle(result_path, {(cam_w, cam_h): compiled_jit}, {'device': Device.DEFAULT})
 
 
 if __name__ == "__main__":
@@ -421,12 +398,6 @@ if __name__ == "__main__":
   model_w, model_h = args.model_size
   output_data = {}
 
-  args.vision_onnx = read_file_chunked_to_disk(args.vision_onnx)
-  args.policy_onnx = read_file_chunked_to_disk(args.policy_onnx)
-  args.off_policy_onnx = read_file_chunked_to_disk(args.off_policy_onnx)
-  args.on_policy_onnx = read_file_chunked_to_disk(args.on_policy_onnx)
-  args.supercombo_onnx = read_file_chunked_to_disk(args.supercombo_onnx)
-
   is_unified_supercombo = False
   if args.model_type == 'supercombo':
     assert args.supercombo_onnx
@@ -449,8 +420,10 @@ if __name__ == "__main__":
                                   frame_copy_size=frame_copy_size)
       warp = stock.make_warp(nv12, model_w, model_h)
       run_model_jit = TinyJit(stock.make_run_model(warp, run_policy, model_metadata, frame_copy_size), prune=True)
-      compiled_jit = compile_jit(run_model_jit, stock.MODELD_INPUTS, make_model_queues, benchmark_runs=args.benchmark_runs, return_loaded=False)
-      output_data[(cam_w, cam_h)] = compiled_jit
+      write_bundle(part_path := f"{args.output}.{cam_w}x{cam_h}", {(cam_w, cam_h): compile_jit(run_model_jit, stock.MODELD_INPUTS, make_model_queues,
+                                                                                               benchmark_runs=args.benchmark_runs)})
+      output_data[(cam_w, cam_h)] = read_parts(part_path)[0][(cam_w, cam_h)]
+      os.remove(part_path)
       gc.collect()
     del model_runner, run_policy
     gc.collect()
@@ -518,18 +491,12 @@ if __name__ == "__main__":
 
       tmp_files.append((cam_w, cam_h, tmp_res.name))
     for cam_w, cam_h, tmp_name in tmp_files:
-      with open(tmp_name, "rb") as f:
-        compiled_jit, dev_name = load_oob(f)
-      output_data[(cam_w, cam_h)] = compiled_jit
-      output_data['input_devices']['warp'] = dev_name
+      warp_parts, warp_extras = read_parts(tmp_name)
+      output_data[(cam_w, cam_h)] = warp_parts[(cam_w, cam_h)]
+      output_data['input_devices']['warp'] = warp_extras['device']
       os.remove(tmp_name)
     output_data['metadata']['warp_dev'] = Device.DEFAULT
 
-  with open(args.output, "wb") as 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)")
-  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)")
+  parts = {key: output_data.pop(key) for key in [key for key, value in output_data.items() if isinstance(value, tuple)]}
+  write_bundle(args.output, parts, output_data)
+  print(f"Saved model to {args.output} ({os.path.getsize(args.output) / 1e6:.2f} MB, {len(parts)} JITs)")

openpilot/sunnypilot/modeld_v2/compile_modeld_v2.py

@@ -1,11 +1,59 @@
 #!/usr/bin/env python3
-import runpy
+import argparse
 import os
+import sys
 
-from openpilot.sunnypilot.modeld_v2.compile_modeld import _patch_ops_serialization
+import numpy as np
+
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, write_bundle
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame, record_kernel_linear, serialize_jit
+
+
+def initialize_check_inputs(arrays: dict):
+  rng = np.random.default_rng(7)
+  for value in arrays.values():
+    value[...] = rng.standard_normal(value.shape) if value.dtype.kind == "f" else rng.integers(0, 2 if value.dtype == np.bool_ else 16, value.shape)
 
 
 if __name__ == "__main__":
-  _patch_ops_serialization()
-  upstream_script = os.path.join(os.environ.get("TINYGRAD_PATH", "tinygrad_repo"), "examples/openpilot/compile_onnx.py")
-  runpy.run_path(upstream_script, run_name="__main__")
+  sys.path.insert(0, os.environ.get("TINYGRAD_PATH", "tinygrad_repo"))
+  from examples.openpilot.compile_onnx import onnx_metadata
+  from examples.openpilot.helpers import allocate_inputs, benchmark
+  from tinygrad import Device, Tensor, TinyJit
+  from tinygrad.dtype import _to_np_dtype
+  from tinygrad.nn.onnx import OnnxRunner
+
+  parser = argparse.ArgumentParser(description="Compile a native-architecture ONNX model into a plain tinygrad artifact")
+  parser.add_argument('onnx')
+  parser.add_argument('output')
+  parser.add_argument('--device-input', action='append', default=[])
+  parser.add_argument('--benchmark-runs', type=int, default=20)
+  args = parser.parse_args()
+
+  runner, (properties, output_shapes) = OnnxRunner(args.onnx), onnx_metadata(args.onnx)
+  device_inputs = runner.graph_inputs.keys() if '*' in args.device_input else args.device_input
+  metadata = {'metadata': properties} | {f'{kind}_shapes': {name: tuple(d if isinstance(d, int) else 0 for d in shape) for name, shape in shapes.items()}
+                                         for kind, shapes in [('input', {name: spec.shape for name, spec in runner.graph_inputs.items()}),
+                                                              ('output', output_shapes)]}
+  input_specs = {name: (tuple(s if isinstance(s, int) else 1 for s in spec.shape), np.dtype(_to_np_dtype(spec.dtype)).str,
+                        Device.DEFAULT if name in device_inputs else 'NPY') for name, spec in runner.graph_inputs.items()}
+
+  def model(inputs):
+    return {name: value.contiguous() for name, value in runner({name: value.to(Device.DEFAULT) for name, value in inputs.items()}).items()}
+  output_specs = {name: (value.shape, np.dtype(_to_np_dtype(value.dtype)).name, Device.DEFAULT) for name, value in model(allocate_inputs(input_specs)).items()}
+
+  @TinyJit(prune=True)
+  def jit(output_buffers, **inputs):
+    Tensor.realize(*(output_buffers[name].assign(value) for name, value in model(inputs).items()))
+
+  check_inputs = allocate_inputs(input_specs, initialize_check_inputs) | {'output_buffers': allocate_inputs(output_specs)}
+  with record_kernel_linear() as recorded:
+    for i in range(2 + args.benchmark_runs):
+      benchmark(jit, cb=lambda t, i=i: print(f"  [{i + 1}/{2 + args.benchmark_runs}] {t * 1e3:.2f} ms"), **check_inputs)
+  assert len(recorded) == 1, f"expected one captured kernel list, got {len(recorded)}"
+
+  frame, _ = capture_frame(jit, check_inputs, 'output_buffers')
+  part = serialize_jit(jit, recorded[0], frame)
+  PlainJit(part[0], memoryview(part[1]))
+  write_bundle(args.output, {'run': part}, {'metadata': metadata, 'input_specs': input_specs, 'output_specs': output_specs})
+  print(f"Saved model to {args.output} ({os.path.getsize(args.output) / 1e6:.2f} MB, {len(part[0]['calls'])} kernels)")

openpilot/sunnypilot/modeld_v2/compile_warp.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import argparse
+import numpy as np
+
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, write_bundle
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame, record_kernel_linear, serialize_jit
+
+
+def main():
+  parser = argparse.ArgumentParser(description="Compile the two-frame yuv420 camera warp into a plain tinygrad artifact")
+  parser.add_argument("--frame", required=True, help="width,height,stride,y_height,uv_height,buffer_size")
+  parser.add_argument("--warp-to", required=True, help="model input WxH")
+  parser.add_argument("--output", required=True)
+  args = parser.parse_args()
+
+  from examples.openpilot.compile_warp import make_frame_prepare, parse_frame, parse_size
+  from tinygrad import Device, Tensor, TinyJit
+  frame = parse_frame(args.frame)
+  warp = make_frame_prepare(frame, *parse_size(args.warp_to))
+  run = TinyJit(lambda input_frame, M_inv: Tensor.stack(*(warp(input_frame[i], M_inv[i]) for i in range(2))), prune=True)
+  rng = np.random.default_rng(42)
+  inputs = {"input_frame": Tensor(rng.integers(0, 256, (2, frame.size), dtype=np.uint8), device=Device.DEFAULT).realize(),
+            "M_inv": Tensor((rng.standard_normal((2, 3, 3)) * 8).astype(np.float32), device=Device.DEFAULT).realize()}
+  with record_kernel_linear() as recorded:
+    run(**inputs)
+    run(**inputs)
+    check, _ = capture_frame(run, inputs)
+  assert len(recorded) == 1, f"expected one captured kernel list, got {len(recorded)}"
+  part = serialize_jit(run, recorded[0], check)
+  PlainJit(*part)
+  specs = {"input_frame": ((2, frame.size), "|u1", Device.DEFAULT), "M_inv": ((2, 3, 3), "<f4", Device.DEFAULT)}
+  write_bundle(args.output, {"run": part}, {"input_specs": specs, "metadata": {}})
+  print(f"Saved warp to {args.output} ({len(part[0]['calls'])} kernels)")
+
+
+if __name__ == "__main__":
+  main()

openpilot/sunnypilot/modeld_v2/convert_upstream_pkl.py

@@ -0,0 +1,214 @@
+#!/usr/bin/env python3
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import argparse
+import dataclasses
+import itertools
+import mmap
+import re
+import struct
+import tempfile
+import numpy as np
+
+from openpilot.sunnypilot.modeld_v2.helpers import ALIGNMENT, Blob, PlainJit, flatten, patch_fetch_fw, write_bundle
+from openpilot.sunnypilot.modeld_v2.plain_artifact_compiler import capture_frame
+
+
+def _root(u):
+  from tinygrad.uop.ops import Ops
+  while u.op in (Ops.AFTER, Ops.BITCAST, Ops.SHRINK, Ops.INDEX):
+    u = u.src[0]
+  return u
+
+
+def _view(u) -> tuple[int, int | None]:
+  from tinygrad.uop.ops import Ops
+  start, size = 0, None
+  while u.op in (Ops.AFTER, Ops.BITCAST, Ops.SHRINK):
+    if u.op is Ops.SHRINK:
+      start, size = start + u.src[1].arg, u.src[2].arg if size is None else size
+    u = u.src[0]
+  return start, size
+
+
+def _address(v):
+  from tinygrad.uop.ops import Ops
+  add = 0
+  while v.op is not Ops.GETADDR:
+    add += v.src[1].arg if v.op is Ops.ADD else 0
+    v = v.src[0]
+  start, size = _view(v.src[0])
+  return _root(v.src[0]), add + start, size
+
+
+def _patched(buf, topo) -> tuple[bytearray, dict]:
+  from tinygrad.uop.ops import Ops
+  blob, rows = bytearray(buf.max_numel() * buf.dtype.itemsize), {}
+  for u in topo:
+    if u.op is not Ops.STORE or _root(u.src[0]) is not buf:
+      continue
+    if (value := u.src[1]).op is Ops.BINARY or value.src[0].op is Ops.BINARY:
+      data = (value if value.op is Ops.BINARY else value.src[0]).arg
+      blob[:len(data)] = data
+    else:
+      view, offsets = u.src[0].src
+      rows |= {_view(view)[0] + o.arg * view.dtype.itemsize: w for o, w in zip(offsets.src, value.src, strict=True)}
+  return blob, rows
+
+
+def _evaluate(u, env):
+  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)
+  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.LOAD:
+    table, index = _root(u.src[0].src[0]), _evaluate(u.src[0].src[1], env)
+    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]):
+    return exec_alu(u.op, u.dtype, values, False)
+  return None
+
+
+def _input_pokes(submit, kernargs) -> dict[int, tuple[int, int]]:
+  from tinygrad.uop.ops import Ops
+  outer = dict(enumerate(submit.src[1:]))
+  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}:
+      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))
+  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))
+    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)
+
+
+def convert(jit, kwargs: dict, frame: dict, target) -> tuple[dict, memoryview]:
+  import tinygrad.runtime.autogen.am.pm4_nv as pm4
+  import tinygrad.runtime.autogen.amd_gpu as amd_gpu
+  from tinygrad.dtype import dtypes
+  from tinygrad.engine.jit import _prepare_jit_inputs
+  from tinygrad.helpers import unwrap
+  from tinygrad.uop.ops import Ops
+  submit = jit.captured._linear.src[0].without_after
+  kernels, topo = submit.arg.aux.kernels, list(jit.captured._linear.toposort(enter_calls=True))
+  allocs = {str(u.tag): u for u in topo if u.op is Ops.ALLOC and u.tag is not None}
+  (commands, patched), kernargs = _patched(allocs["cmdbuf_compute_0"], topo), allocs["kernargs_compute_0"]
+  argument_words, pokes = _patched(kernargs, topo)[1], _input_pokes(submit, kernargs)
+  sh = {name: getattr(amd_gpu, f"reg{name}") + amd_gpu.GC_BASE__INST0_SEG0 - pm4.PACKET3_SET_SH_REG_START
+        for name in ("COMPUTE_PGM_LO", "COMPUTE_USER_DATA_0", "COMPUTE_NUM_THREAD_X")} if hasattr(amd_gpu, "regCOMPUTE_PGM_LO") else \
+       {"COMPUTE_PGM_LO": 0x20c, "COMPUTE_USER_DATA_0": 0x240, "COMPUTE_NUM_THREAD_X": 0x207}
+
+  dispatches, registers, offset = [], {}, 0
+  while offset < len(commands):
+    word = lambda o: patched.get(o, struct.unpack_from("<I", commands, o)[0])  # noqa: E731
+    if not isinstance(header := word(offset), int) or header >> 30 != 3:
+      offset += 4
+      continue
+    count, opcode = ((header >> 16) & 0x3fff) + 1, (header >> 8) & 0xff
+    body = [word(offset + 4 + 4 * i) for i in range(count)]
+    if opcode == pm4.PACKET3_SET_SH_REG:
+      registers |= {body[0] + i: v for i, v in enumerate(body[1:])}
+    elif opcode == pm4.PACKET3_DISPATCH_DIRECT:
+      local = [registers[sh["COMPUTE_NUM_THREAD_X"] + i] for i in range(3)]
+      dispatches.append((registers[sh["COMPUTE_PGM_LO"]], registers[sh["COMPUTE_USER_DATA_0"]], local, body[:3]))
+    offset += 4 + 4 * count
+  assert len(dispatches) == len(kernels), f"decoded {len(dispatches)} dispatches, the JIT has {len(kernels)} kernels"
+
+  sink = tempfile.TemporaryFile(dir=".")
+  blob, buffers, programs = Blob(sink), {}, {}
+  u8 = repr(dtypes.uint8)
+
+  def store(array) -> dict:
+    array = np.ascontiguousarray(array)
+    return {"data": blob.add(array.tobytes()), "dtype": array.dtype.str, "shape": list(array.shape)}
+
+  calls, written, needs_init = [], set(), set()
+  for (program, arguments, local, grid), kernel in zip(dispatches, kernels, strict=True):
+    image_alloc, entry, _ = _address(program)
+    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))
+    start, size = _view(arguments.src[0])
+    args = []
+    for slot in range(start, start + unwrap(size), 8):
+      if slot in pokes:
+        assert pokes[slot][1] == 0, "the plain format passes whole inputs, this kernel takes a view into one"
+        args.append({"input": pokes[slot][0]})
+      elif slot in argument_words:
+        base, at, nbytes = _address(argument_words[slot])
+        args.append({"buffer": buffers.setdefault(base, len(buffers)), "offset": at, "nbytes": nbytes or base.arg.buffer.nbytes, "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),
+                  "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}
+    written |= {args[i]["buffer"] for i in outs if "buffer" in args[i]}
+
+  names = list(jit.captured.expected_names)
+  inputs = [{"slot": slot, "name": names[slot] if slot < len(names) else None, "dtype": repr(u.dtype), "numel": int(u.max_numel()), "device": u.device}
+            for slot, u in enumerate(_prepare_jit_inputs((), kwargs)[0])]
+  header = {"input_names": names, "calls": calls, "returns": {"kind": "NoneType", "tensors": []}, "inputs": inputs,
+            "buffers": [{"nbytes": b.arg.buffer.nbytes, "device": b.arg.buffer.device, "dtype": u8}
+                        | ({"init": blob.add(b.arg.buffer.ensure_allocated().as_memoryview())} if i in needs_init else {}) for b, i in buffers.items()],
+            "check": {"inputs": [[path, store(array), device] for path, array, device in frame["inputs"]],
+                      "outputs": [store(array) for array in frame["outputs"]], "output_container": frame["output_container"]}}
+  with sink:
+    sink.flush()
+    return header, memoryview(mmap.mmap(sink.fileno(), 0, access=mmap.ACCESS_READ))
+
+
+def main():
+  parser = argparse.ArgumentParser(description="Convert commaai's precompiled chestnut driving model into a plain tinygrad artifact")
+  parser.add_argument("pkl", help="big_driving_tinygrad.pkl")
+  parser.add_argument("output")
+  args = parser.parse_args()
+
+  patch_fetch_fw()
+  from examples.openpilot.helpers import allocate_inputs, load_pickle
+  from tinygrad import Device, Tensor
+  from tinygrad.engine.realize import lower_and_compile
+  artifact = load_pickle(args.pkl, out_of_band=True)
+  artifact["run"].captured._linear = lower_and_compile(artifact["run"].captured._linear)
+  rng = np.random.default_rng(7)
+
+  def initialize(arrays: dict):
+    for value in arrays.values():
+      value[...] = rng.standard_normal(value.shape) if value.dtype.kind == "f" else rng.integers(0, 2 if value.dtype == np.bool_ else 16, value.shape)
+  kwargs = allocate_inputs(artifact["input_specs"], initialize) | {"output_buffers": allocate_inputs(artifact["output_specs"])}
+  frame, _ = capture_frame(artifact["run"], kwargs, "output_buffers")
+  part = convert(artifact["run"], kwargs, frame, Device[Device.DEFAULT].renderer.target)
+
+  plain = PlainJit(*part)  # replays the frame, within tolerance
+  replay = {path[0]: Tensor(array, device=device).realize() for path, array, device in frame["inputs"] if len(path) == 1}
+  replay["output_buffers"] = {path[1]: Tensor(array, device=device).realize() for path, array, device in frame["inputs"] if len(path) == 2}
+  flatten(plain(**replay))
+  for index, (got, expected) in enumerate(zip([t.numpy() for t in replay["output_buffers"].values()], frame["outputs"], strict=True)):
+    assert np.array_equal(got, expected), f"output {index} of the plain artifact differs from the pickle"
+  write_bundle(args.output, {"run": part}, {key: artifact[key] for key in ("metadata", "input_specs", "output_specs")})
+  print(f"converted {len(part[0]['calls'])} kernels, bit-exact with the pickle: {args.output}")
+
+
+if __name__ == "__main__":
+  main()

openpilot/sunnypilot/modeld_v2/frame_fusion.py

@@ -0,0 +1,69 @@
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import bisect
+import itertools
+import math
+from functools import partial
+
+from openpilot.common.basedir import BASEDIR
+from openpilot.sunnypilot.modeld_v2.helpers import _dtype, read_parts
+
+
+def pack_layout(shapes: dict[str, tuple], frame_size: int) -> tuple[dict[str, int], int]:
+  offsets, offset = {}, 0
+  for name, shape in ({"tfm": (2, 3, 3)} | shapes).items():
+    offsets[name], offset = offset, offset + -(-math.prod(shape) * 4 // 128) * 128
+  return offsets | {"frames": offset}, offset + 2 * frame_size
+
+
+class _Blobs:  # the model's and the warp's blobs as one address space, for a part whose ranges span both files
+  def __init__(self, *blobs):
+    self.blobs, self.starts = blobs, list(itertools.accumulate([0] + [len(blob) for blob in blobs[:-1]]))
+
+  def __getitem__(self, span: slice):
+    index = bisect.bisect_right(self.starts, span.start) - 1
+    return self.blobs[index][span.start - self.starts[index]:span.stop - self.starts[index]]
+
+
+def fused_with_warp(cam_w: int, cam_h: int, chestnut: bool):  # load_model's extend: a native model runs upload, warp and model as one submit
+  return partial(frame_part, f"{BASEDIR}/openpilot/sunnypilot/modeld_v2/models/{'big_' if chestnut else ''}driving_warp_{cam_w}x{cam_h}_tinygrad.pkl")
+
+
+def frame_part(warp_path: str, parts: dict, extras: dict) -> dict:
+  if "input_specs" not in extras:  # a legacy model warps in its own parts
+    return {}
+  (model, model_blob), (warp, warp_blob) = parts["run"], read_parts(warp_path)[0]["run"]
+  states = sorted(name for name in extras["input_specs"] if f"next_{name}" in extras["metadata"]["output_shapes"])
+  host = {name: shape for name, (shape, _, _) in extras["input_specs"].items() if name not in states and name != "new_img"}
+  frame_size = next(s["numel"] for s in warp["inputs"] if s["name"] == "input_frame") // 2
+  offsets, packed = pack_layout(host, frame_size)
+  shift, upload, u8 = len(model_blob), len(model["buffers"]) + len(warp["buffers"]), repr(_dtype("dtypes.uint8"))
+
+  def view(name: str, nbytes: int, dtype: str) -> dict:
+    return {"buffer": upload, "offset": offsets[name], "nbytes": nbytes, "dtype": dtype}
+  outputs = sorted(s["slot"] for s in model["inputs"] if s["name"] is None)
+  names = sorted(["host", *states])
+  slots = {s["slot"]: names.index(s["name"]) for s in model["inputs"] if s["name"] in states} | \
+          {slot: len(names) + i for i, slot in enumerate(outputs)}
+  returned = warp["returns"]["tensors"][0]
+  model_inputs = {s["slot"]: {"input": slots[s["slot"]]} if s["slot"] in slots else
+                  {k: v for k, v in returned.items() if k != "shape"} | {"buffer": returned["buffer"] + len(model["buffers"])} if s["name"] == "new_img" else
+                  view(s["name"], s["numel"] * _dtype(s["dtype"]).itemsize, s["dtype"]) for s in model["inputs"]}
+  warp_inputs = {s["slot"]: view("frames" if s["name"] == "input_frame" else "tfm", s["numel"] * _dtype(s["dtype"]).itemsize, s["dtype"])
+                 for s in warp["inputs"]}
+
+  def remap(calls: list, inputs: dict, buffers: int, lib: int) -> list:
+    return [call | ({"lib": [call["lib"][0] + lib, call["lib"][1] + lib]} if "lib" in call else {}) |
+            {"args": [dict(inputs[a["input"]]) if "input" in a else a | {"buffer": a["buffer"] + buffers} for a in call["args"]]} for call in calls]
+  header = {"input_names": names, "returns": {"kind": "NoneType", "tensors": []}, "check": None,
+            "inputs": [{"slot": names.index("host"), "name": "host", "dtype": u8, "numel": packed, "device": "NPY"}] +
+                      [s | {"slot": slots[s["slot"]]} for s in model["inputs"] if s["slot"] in slots],
+            "buffers": model["buffers"] + [b | ({"init": [b["init"][0] + shift, b["init"][1] + shift]} if "init" in b else {}) for b in warp["buffers"]] +
+                       [{"nbytes": packed, "device": model["inputs"][0]["device"], "dtype": u8}],
+            "calls": [{"kind": "copy", "args": [{"buffer": upload, "offset": 0, "nbytes": packed, "dtype": u8}, {"input": names.index("host")}]}] +
+                     remap(warp["calls"], warp_inputs, len(model["buffers"]), shift) + remap(model["calls"], model_inputs, 0, 0)}
+  return {"run": (header, _Blobs(model_blob, warp_blob))}  # replaces the model's run: one set of buffers, no separate warp

openpilot/sunnypilot/modeld_v2/helpers.py

@@ -5,14 +5,25 @@ This file is part of sunnypilot and is licensed under the MIT License.
 See the LICENSE.md file in the root directory for more details.
 """
 
+import bisect
+import builtins
+import contextlib
+import dataclasses
+import functools
+import hashlib
 import io
-import struct
+import itertools
+import json
+import math
+import mmap
+import os
 import pickle
-import inspect
-import importlib
-import tempfile
-import shutil
-import enum
+import struct
+
+MAGIC = b"SPTGPLN1"
+ALIGNMENT = 256
+METADATA_TYPES = ("slice", "tuple", "list", "dict", "set", "frozenset", "complex")
+RUNTIME_ENVIRONMENT = ("GMMU", "AMD_AQL", "AMD_DISABLE_SDMA", "GRAPH_ONE_KERNEL")
 
 
 def _patch_system_flock_acquire():
@@ -32,86 +43,289 @@ def _patch_system_flock_acquire():
     pass
 _patch_system_flock_acquire()
 
-def _pad_args(func, args, kwargs):
+
+def patch_fetch_fw() -> None:
   try:
-    sig = inspect.signature(func)
-  except Exception:
-    return args, kwargs
-  params = list(sig.parameters.values())
-  if inspect.isfunction(func) and params and params[0].name in ('cls', 'self'):
-    params = params[1:]
-
-  new_args = list(args)
-  has_varargs = any(p.kind == inspect.Parameter.VAR_POSITIONAL for p in params)
-  if len(new_args) > len(params) and not has_varargs:
-    new_args = new_args[:len(params)]
-
-  for i in range(len(new_args), len(params)):
-    param = params[i]
-    if param.kind in (inspect.Parameter.VAR_POSITIONAL, inspect.Parameter.VAR_KEYWORD):
-      continue
-    val = param.default if param.default is not inspect.Parameter.empty else None
-    new_args.append(val)
-  return new_args, kwargs
+    import zstandard
+  except ImportError:
+    return
+  zstandard.decompress = lambda data, *_: zstandard.ZstdDecompressor().stream_reader(data).read()  # ty: ignore[invalid-assignment]
 
 
-def _dynamic_factory(real_class):
-  if isinstance(real_class, type) and issubclass(real_class, enum.Enum):
-    return real_class
-
-  def factory(*args, **kwargs):
-    try:
-      return real_class(*args, **kwargs)
-    except TypeError:
-      new_args, new_kwargs = _pad_args(real_class, args, kwargs)
-      return real_class(*new_args, **new_kwargs)
-
-  class DynamicMeta(type(real_class)):
-    def __call__(cls, *args, **kwargs):
-      return factory(*args, **kwargs)
-
-  class DynamicProxy(real_class, metaclass=DynamicMeta):
-    __slots__ = ()
-
-    def __new__(cls, *args, **kwargs):
-      return factory(*args, **kwargs)
-
-  DynamicProxy.__name__ = real_class.__name__
-  DynamicProxy.__module__ = real_class.__module__
-  return DynamicProxy
+class ArtifactMismatch(RuntimeError):
+  pass
 
 
-class DynamicTinygradUnpickler(pickle.Unpickler):
+class Blob:
+  def __init__(self, sink=None):
+    self.chunks: list = []
+    self.locations: dict[bytes, list[int]] = {}
+    self.size, self.write = 0, sink.write if sink else self.chunks.append
+
+  def add(self, data) -> list[int]:
+    if (key := hashlib.blake2b(data, digest_size=16).digest()) not in self.locations:
+      start = self.size + -self.size % ALIGNMENT
+      self.write(bytes(start - self.size)), self.write(data)
+      self.locations[key], self.size = [start, start + memoryview(data).nbytes], start + memoryview(data).nbytes
+    return self.locations[key]
+
+
+class _MetadataUnpickler(pickle.Unpickler):
   def find_class(self, module, name):
-    real_class = getattr(importlib.import_module(module), name)
-    if module.startswith("tinygrad"):
-      return _dynamic_factory(real_class)
-    return real_class
+    if module == "builtins" and name in METADATA_TYPES:
+      return getattr(builtins, name)
+    raise pickle.UnpicklingError(f"{module}.{name} is not allowed in model metadata")
 
 
-def load_oob(f):
-  opcodes = f.read(struct.unpack('<q', f.read(8))[0])
-  def buffers():
-    while (h := f.read(8)):
-      pb = pickle.PickleBuffer(bytearray(struct.unpack('<q', h)[0]))
-      f.readinto(pb)
-      yield pb
-  return DynamicTinygradUnpickler(io.BytesIO(opcodes), buffers=buffers()).load()
+def flatten(value) -> list:
+  if isinstance(value, (tuple, list)):
+    return [tensor for item in value for tensor in flatten(item)]
+  return [] if value is None else [value]
 
 
-def dump_oob(obj, f):
-  with tempfile.TemporaryFile(dir=".") as tmp:
-    def buffer_cb(buffer):
-      data = buffer.raw()
-      tmp.write(struct.pack('<q', len(data)))
-      tmp.write(data)
-      buffer.release()
-      return False
+def _relocate(value, part_blob: memoryview, blob: Blob):
+  if isinstance(value, dict):
+    return {k: blob.add(part_blob[slice(*v)]) if k in ("lib", "init", "data") else _relocate(v, part_blob, blob) for k, v in value.items()}
+  return [_relocate(v, part_blob, blob) for v in value] if isinstance(value, list) else value
 
-    stream = io.BytesIO()
-    pickle.Pickler(stream, protocol=5, buffer_callback=buffer_cb).dump(obj)
-    opcodes = stream.getvalue()
-    f.write(struct.pack('<q', len(opcodes)))
-    f.write(opcodes)
-    tmp.seek(0)
-    shutil.copyfileobj(tmp, f)
+
+def write_bundle(path: str, parts: dict, extras: dict | None = None):
+  blob = Blob()
+  headers = {key: _relocate(part_header, memoryview(part_blob), blob) for key, (part_header, part_blob) in parts.items()}
+  extras_location = blob.add(pickle.dumps(extras or {}, protocol=4))
+  entries = [{"key": list(key) if isinstance(key, tuple) else key, "blob": [0, blob.size], "header": header} for key, header in headers.items()]
+  header = json.dumps({"parts": entries, "extras": extras_location}).encode()
+  with open(path, "wb") as file:
+    file.write(MAGIC + struct.pack("<Q", len(header)) + header + bytes(-(len(MAGIC) + 8 + len(header)) % ALIGNMENT))
+    file.writelines(blob.chunks)
+
+
+def read_parts(path: str) -> tuple[dict, dict]:
+  with open(path, "rb") as file:
+    data = memoryview(mmap.mmap(file.fileno(), 0, access=mmap.ACCESS_COPY))
+  if bytes(data[:len(MAGIC)]) != MAGIC:
+    raise ValueError(f"{path} is not a plain tinygrad model, recompile it with compile_modeld.py")
+  start = len(MAGIC) + 8 + struct.unpack_from("<Q", data, len(MAGIC))[0]
+  header, blob = json.loads(bytes(data[len(MAGIC) + 8:start])), data[start + -start % ALIGNMENT:]
+  parts = {tuple(e["key"]) if isinstance(e["key"], list) else e["key"]: (e["header"], blob[slice(*e["blob"])]) for e in header["parts"]}
+  return parts, _MetadataUnpickler(io.BytesIO(bytes(blob[slice(*header["extras"])]))).load()
+
+
+def _extended(path: str, extend) -> tuple[dict, dict, str]:
+  parts, extras = read_parts(path)
+  added = extend(parts, extras) if extend else {}
+  return parts | added, extras, _parts_digest(added) if added else ""
+
+
+def load_model(path: str, device_map: dict | None = None, verify: bool = True, cache: bool = True, extend=None) -> dict:
+  parts, extras, extra = _extended(path, extend)
+  if cache and (cached := _read_cache(path, parts, device_map, extra)) is not None:
+    return extras | cached
+  shared: dict = {}
+  jits = {key: PlainJit(header, blob, device_map, verify, shared) for key, (header, blob) in parts.items()}
+  if cache:
+    _write_cache(path, jits, device_map, extra)
+  return extras | jits
+
+
+@functools.cache
+def tinygrad_identity() -> str:
+  import tinygrad
+  root, digest = os.path.dirname(tinygrad.__file__), hashlib.sha256()
+  for directory, _, files in sorted(os.walk(root)):
+    for name in sorted(file for file in files if file.endswith(".py")):
+      with open(path := os.path.join(directory, name), "rb") as file:
+        digest.update(f"{os.path.relpath(path, root)}:".encode() + file.read())
+  return digest.hexdigest()
+
+
+def _parts_digest(parts: dict) -> str:
+  digest = hashlib.sha256()
+  for _, (header, blob) in sorted(parts.items(), key=lambda item: repr(item[0])):
+    digest.update(json.dumps(header, sort_keys=True).encode())
+    for call in header["calls"]:
+      if call["kind"] == "program":
+        digest.update(blob[slice(*call["lib"])])
+  return digest.hexdigest()
+
+
+def _cache_key(path: str, device_map: dict | None, extra: str = "") -> bytes:
+  stat = os.stat(path)
+  return json.dumps({"format": MAGIC.decode(), "tinygrad": tinygrad_identity(), "size": stat.st_size, "mtime_ns": stat.st_mtime_ns,
+                     "extra": extra, "device_map": sorted((device_map or {}).items()),
+                     "env": [os.environ.get(name, "") for name in RUNTIME_ENVIRONMENT]}).encode()
+
+
+def _cache_header_matches(file, path: str, device_map: dict | None, extra: str = "") -> bool:
+  return file.read(struct.unpack("<Q", file.read(8))[0]) == _cache_key(path, device_map, extra)
+
+
+def cache_is_fresh(path: str, device_map: dict | None = None) -> bool: # entry point for warmup proc
+  try:
+    with open(f"{path}.cache", "rb") as file:
+      return _cache_header_matches(file, path, device_map)
+  except (OSError, struct.error):
+    return False
+
+
+def _read_cache(path: str, parts: dict, device_map: dict | None, extra: str = "") -> dict | None:
+  try:
+    with open(f"{path}.cache", "rb") as file:
+      if not _cache_header_matches(file, path, device_map, extra):
+        return None
+      shared: dict = {}
+      buffers = {repr(key): _allocate_buffers(header, blob, device_map, shared) for key, (header, blob) in parts.items()}
+      return type("_Unpickler", (pickle.Unpickler,), {"persistent_load": lambda self, pid: buffers[pid[0]][pid[1]]})(file).load()
+  except Exception:
+    return None
+
+
+def _write_cache(path: str, jits: dict, device_map: dict | None, extra: str = "") -> None:
+  from tinygrad.device import Buffer
+  owners = {id(buffer): (repr(key), index) for key, jit in jits.items() for index, buffer in enumerate(jit.buffers)}
+
+  class _Pickler(pickle.Pickler):
+    def persistent_id(self, obj):
+      return owners.get(id(obj)) if isinstance(obj, Buffer) else None
+
+  key, tmp_path = _cache_key(path, device_map, extra), f"{path}.cache.{os.getpid()}.tmp"
+  with contextlib.suppress(Exception):
+    with open(tmp_path, "wb") as file:
+      file.write(struct.pack("<Q", len(key)) + key)
+      _Pickler(file, protocol=5).dump({name: _Jit(jit.captured, jit.input_names, jit.buffers) for name, jit in jits.items()})
+    os.replace(tmp_path, f"{path}.cache")
+
+
+# Everything below this comment block calls tinygrad internals, so after a tinygrad bump this
+# is the only code to update, and models are not recompiled. How to do this when discountchubbs is not available:
+#   1. run modeld_v2/tests/test_plain_artifact.py on the new tinygrad and fix each failing call against the new API: Buffer,
+#      UOp.from_buffer/param (ParamArg), ProgramInfo, Target, KernelInfo, compile_linear, CapturedJit and
+#      _prepare_jit_inputs in tinygrad/engine/jit.py
+#   2. load a real model with verify=True: ArtifactMismatch means the rebuilt graph still differs from the one that was compiled
+# The boot cache is keyed on the tinygrad source, so it rebuilds itself. Only a change in how the device runs kernel binaries needs a recompile.
+# Honestly, I would just have your clanker try it because you likely have zero idea what you are looking at lol..
+
+def _dtype(text: str):
+  from tinygrad.dtype import dtypes
+  return eval(text, {"__builtins__": {}}, {"dtypes": dtypes})
+
+
+def _device(recorded: str, device_map: dict | None) -> str:
+  from tinygrad import Device
+  device_map = device_map or {}
+  return Device.canonicalize(device_map.get(recorded, device_map.get(recorded.split(":", 1)[0], recorded)))
+
+
+def _upload(buffer, data: memoryview, chunk: int = 16 << 20) -> None:
+  from tinygrad.device import Buffer
+  source, chunk = Buffer("NPY", len(data), opaque=memoryview(bytearray(data)) if data.readonly else data), min(chunk, len(data))
+  for offset in sorted({*range(0, len(data) - chunk, chunk), len(data) - chunk}):
+    buffer.view(chunk, offset).ensure_allocated().copy_from(source.view(chunk, offset).ensure_allocated())
+
+
+def _allocate_buffers(header: dict, blob: memoryview, device_map: dict | None, shared: dict) -> list:
+  import numpy as np
+  from tinygrad.device import Buffer
+  written = {call["args"][i]["buffer"] for call in header["calls"] for i in (call["outs"] if call["kind"] == "program" else (0,))
+             if "buffer" in call["args"][i]}
+  entries = [(_device(entry["device"], device_map), blob[slice(*entry["init"])] if "init" in entry else None, entry) for entry in header["buffers"]]
+  keys = [(device, entry["dtype"], np.frombuffer(data, np.uint8).ctypes.data, len(data)) if data and index not in written else object()
+          for index, (device, data, entry) in enumerate(entries)]
+  for device in {key[0] for key in keys if isinstance(key, tuple) and key not in shared}:
+    if shared.setdefault(("mapped", device), Buffer(device, 1).ensure_allocated().get_storage().host is not None):
+      continue
+    spans = [(entry["init"], entry["nbytes"], key) for (_, _, entry), key in zip(entries, keys, strict=True)
+             if isinstance(key, tuple) and key[0] == device and key not in shared]
+    files = getattr(blob, "starts", [0])  # a part read from several files (frame_fusion's) is uploaded a block per file
+    for _, group in itertools.groupby(sorted(spans, key=lambda span: span[0][0]), key=lambda span: bisect.bisect_right(files, span[0][0])):
+      group = list(group)
+      start, end = min(span[0] for span, _, _ in group), max(span[1] for span, _, _ in group)
+      _upload(block := Buffer(device, end - start).ensure_allocated(), blob[start:end])
+      shared |= {key: block.view(nbytes, span[0] - start).ensure_allocated() for span, nbytes, key in group}
+  buffers = []
+  for (device, data, entry), key in zip(entries, keys, strict=True):
+    if key not in shared:
+      shared[key] = Buffer(device, entry["nbytes"], initial_value=pickle.PickleBuffer(data) if data else None).ensure_allocated()
+    buffers.append(shared[key])
+  return buffers
+
+
+def _program_uop(call: dict, lib: bytes):
+  from tinygrad.helpers import Target
+  from tinygrad.uop.ops import KernelInfo, Ops, ProgramInfo, UOp
+  parameters = []
+  for name, compact_slot, dtype, shape in call["signature"]:
+    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=dataclasses.replace(parameter.arg, name=name, image=image)))
+  info = ProgramInfo(global_size=tuple(call["global_size"]), local_size=tuple(call["local_size"]), vars=(), globals=tuple(call["globals"]),
+                     outs=tuple(call["outs"]), ins=tuple(call["ins"]), target=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)
+
+
+class _Jit:
+  def __init__(self, captured, input_names: list, buffers: list):
+    self.captured, self.input_names, self.buffers = captured, input_names, buffers
+
+  def __call__(self, *args, **kwargs):
+    import tinygrad.engine.jit as jit
+    input_uops, var_vals, names, _ = jit._prepare_jit_inputs(args, kwargs)
+    if names != self.input_names:
+      raise ValueError(f"expected inputs {self.input_names}, got {names}")
+    return self.captured(input_uops, var_vals)
+
+
+class PlainJit(_Jit):
+  def __init__(self, header: dict, blob: memoryview, device_map: dict | None = None, verify: bool = True, shared: dict | None = None):
+    import numpy as np
+    import tinygrad.engine.jit as jit
+    from tinygrad import Tensor
+    from tinygrad.device import Buffer
+    from tinygrad.dtype import dtypes
+    from tinygrad.uop.ops import Ops, UOp
+
+    inputs = {spec["slot"]: spec for spec in header["inputs"]}
+    buffers = _allocate_buffers(header, blob, device_map, {} if shared is None else shared)
+    bases = [UOp.from_buffer(buffer, _dtype(entry["dtype"])) for buffer, entry in zip(buffers, header["buffers"], strict=True)]
+
+    def argument(described: dict):
+      if "input" in described:
+        spec = inputs[described["input"]]
+        return UOp.param(described["input"], _dtype(spec["dtype"]), spec["numel"], _device(spec["device"], device_map))
+      base, dtype, offset, nbytes = bases[described["buffer"]], _dtype(described["dtype"]), described["offset"], described["nbytes"]
+      if offset == 0 and nbytes == header["buffers"][described["buffer"]]["nbytes"] and dtype == base.dtype:
+        return base
+      return (base if base.dtype.itemsize == 1 else base.bitcast(dtypes.uint8)).shrink(((offset, offset + nbytes),)).bitcast(dtype)
+
+    calls = []
+    for call in header["calls"]:
+      arguments = [argument(described) for described in call["args"]]
+      if call["kind"] == "program":
+        calls.append(_program_uop(call, bytes(blob[slice(*call["lib"])])).call(*arguments))
+      else:
+        calls.append(arguments[0].store_call(arguments[1]))
+    # stand-ins for JIT inputs: copy staging to GPU is decided from the input buffers, the real ones bind per call
+    stand_ins = {slot: UOp.from_buffer(Buffer(_device(spec["device"], device_map), spec["numel"] * _dtype(spec["dtype"]).itemsize).ensure_allocated(),
+                                       _dtype(spec["dtype"])) for slot, spec in inputs.items()}
+    input_uops = tuple(stand_ins.get(slot, next(iter(stand_ins.values()), None)) for slot in range(max(stand_ins, default=-1) + 1))
+    linear = jit.compile_linear(UOp(Ops.LINEAR, src=tuple(calls)), beam=0, input_uops=input_uops, cache=False)
+
+    returned = [Tensor(argument(t)).reshape(tuple(t["shape"])) for t in header["returns"]["tensors"]]
+    ret = {"NoneType": None, "Tensor": returned[0] if returned else None, "tuple": tuple(returned), "list": returned}[header["returns"]["kind"]]
+    super().__init__(jit.CapturedJit(ret, linear, header["input_names"], []), header["input_names"], buffers)
+
+    if not verify or not (check := header.get("check")):
+      return
+    def array(stored):
+      return np.frombuffer(blob[slice(*stored["data"])], dtype=np.dtype(stored["dtype"])).reshape(stored["shape"]).copy()
+    kwargs: dict = {}
+    for path, stored, device in check["inputs"]:
+      tensor = Tensor(array(stored), device=_device(device, device_map)).realize()
+      kwargs[path[0]] = tensor if len(path) == 1 else {**kwargs.get(path[0], {}), path[1]: tensor}
+    result = self(**kwargs)
+    outputs = kwargs[check["output_container"]].values() if check["output_container"] else flatten(result)
+    for index, (got, stored) in enumerate(zip([t.numpy() for t in outputs], check["outputs"], strict=True)):
+      expected = array(stored)
+      if got.shape != expected.shape or not np.allclose(got, expected, rtol=1e-3, atol=1e-3, equal_nan=True):
+        raise ArtifactMismatch(f"output {index} differs from the compile-time reference frame; recompile the model")

openpilot/sunnypilot/modeld_v2/model_adapters.py

@@ -11,20 +11,21 @@ import pickle
 import numpy as np
 
 from openpilot.common.basedir import BASEDIR
+from openpilot.sunnypilot.modeld_v2.frame_fusion import pack_layout
+from openpilot.sunnypilot.modeld_v2.helpers import load_model
 from openpilot.sunnypilot.modeld_v2.compile_modeld import (POLICY_INPUTS, derive_frame_skip,
                                                            make_split_input_queues, make_supercombo_input_queues)
 from openpilot.sunnypilot.modeld_v2.stock_dependencies import make_input_queues as stock_make_input_queues, nv12_copy_size
 from openpilot.system.camerad.cameras.nv12_info import get_nv12_info
 from tinygrad.device import Buffer
-from tinygrad.dtype import DType, dtypes
-from tinygrad.helpers import round_up
+from tinygrad.dtype import DType
 from tinygrad.tensor import Tensor
 from tinygrad.uop.ops import UOp
 
 
 def input_view(buffer: Buffer, shape: tuple[int, ...], dtype: DType, offset: int) -> Tensor:
-  view = buffer.view(math.prod(shape), dtype, offset).ensure_allocated()
-  return Tensor(UOp.from_buffer(view)).reshape(shape)
+  view = buffer.view(math.prod(shape) * dtype.itemsize, offset).ensure_allocated()
+  return Tensor(UOp.from_buffer(view, dtype)).reshape(shape)
 
 
 class BaseModelAdapter:
@@ -64,16 +65,10 @@ class BaseModelAdapter:
       self.warp_frame_size = self.nv12_info[3]
       return self.jits[(self.cam_w, self.cam_h)]
 
-    warp_dir = Path(BASEDIR) / "openpilot/sunnypilot/modeld_v2/models"
     warp_name = f'{"big_" if self.chestnut else ""}driving_warp_{self.cam_w}x{self.cam_h}_tinygrad.pkl'
-    with open(warp_dir / warp_name, 'rb') as f:
-      warp_data = pickle.load(f)
-      run_warp = warp_data['run']
-      if 'input_specs' in warp_data and 'input_frame' in warp_data['input_specs']:
-        self.warp_frame_size = warp_data['input_specs']['input_frame'][0][1]
-      else:
-        self.warp_frame_size = self.nv12_info[3] if self.chestnut else self.frame_copy_size
-      return run_warp
+    warp = load_model(str(Path(BASEDIR) / "openpilot/sunnypilot/modeld_v2/models" / warp_name), cache=False)
+    self.warp_frame_size = warp['input_specs']['input_frame'][0][1]
+    return warp['run']
 
 
 class LegacyModelAdapter(BaseModelAdapter):
@@ -172,27 +167,19 @@ class NativeTinygradAdapter(BaseModelAdapter):
     self._vision_input_names = [k for k in self.input_shapes_orig if 'img' in k]
     self.vision_output_slices = pickle.loads(codecs.decode(self.jits['metadata']['metadata']['output_slices'].encode(), 'base64'))
 
-    self.run_warp = self._load_warp()
+    self.warp_frame_size = self.frame_copy_size
     self._init_common()
     self.run_model = self.jits['run']
 
-    warp_frame_size = getattr(self, 'warp_frame_size', self.nv12_info[3])
     self.input_queues = {name: Tensor(np.zeros(shape, dtype=dtype), device=self.model_device).realize()
                          for name, (shape, dtype) in self.input_shapes.items() if name in self.state_pairs}
-    shapes = {'tfm': (2, 3, 3)} | {name: shape for name, (shape, _) in self.input_shapes.items()
-                                   if name not in self.state_pairs and name != 'new_img'}
-    npy_size = sum(round_up(math.prod(shape) * 4, 128) for shape in shapes.values())
-    self.packed_input = np.zeros(npy_size + 2 * warp_frame_size, dtype=np.uint8)
-    self.input_host = Tensor(self.packed_input, device='NPY')._buffer()
-    self.input_device = Tensor(self.packed_input, device=self.model_device)._buffer()
-    self.numpy_inputs = {}
-    offset = 0
-    for name, shape in shapes.items():
-      self.numpy_inputs[name] = np.ndarray(shape, dtype=np.float32, buffer=self.packed_input, offset=offset)
-      self.input_queues[name] = input_view(self.input_device, shape, dtypes.float32, offset)
-      offset += round_up(self.numpy_inputs[name].nbytes, 128)
-    self.frames = self.packed_input[npy_size:].reshape(2, warp_frame_size)
-    self.warp_inputs = {'input_frame': input_view(self.input_device, self.frames.shape, dtypes.uint8, npy_size), 'M_inv': self.input_queues.pop('tfm')}
+    shapes = {name: shape for name, (shape, _) in self.input_shapes.items() if name not in self.state_pairs and name != 'new_img'}
+    offsets, packed = pack_layout(shapes, self.warp_frame_size)
+    self.packed_input = np.zeros(packed, dtype=np.uint8)
+    self.host = Tensor(self.packed_input, device='NPY')
+    self.numpy_inputs = {name: np.ndarray(shape, dtype=np.float32, buffer=self.packed_input, offset=offsets[name])
+                         for name, shape in ({'tfm': (2, 3, 3)} | shapes).items()}
+    self.frames = self.packed_input[offsets['frames']:].reshape(2, self.warp_frame_size)
 
     self.outputs = {name: Tensor(np.zeros(shape, dtype=dtype), device=device).realize() for name, (shape, dtype, device) in self.jits['output_specs'].items()}
     for name, next_name in self.state_pairs.items():
@@ -211,9 +198,7 @@ class NativeTinygradAdapter(BaseModelAdapter):
       self.input_queues[key].assign(0).realize()
 
   def run(self):
-    self.input_device.copy_from(self.input_host)
-    self.input_queues['new_img'] = self.run_warp(**self.warp_inputs)
-    self.run_model(output_buffers=self.outputs, **self.input_queues)
+    self.run_model(host=self.host, **self.input_queues, output_buffers=self.outputs)
     return self.outputs['outputs']
 
 

openpilot/sunnypilot/modeld_v2/modeld.py

@@ -24,7 +24,6 @@ from openpilot.cereal.messaging import PubMaster, SubMaster
 from openpilot.cereal.visionipc import VisionStreamType
 from msgq.visionipc import VisionIpcClient, VisionBuf
 from opendbc.car.car_helpers import get_demo_car_params
-from openpilot.common.file_chunker import open_file_chunked
 from openpilot.common.swaglog import cloudlog
 from openpilot.common.params import Params
 from openpilot.common.filter_simple import FirstOrderFilter
@@ -42,7 +41,8 @@ from openpilot.sunnypilot.modeld_v2.meta_helper import load_meta_constants
 from openpilot.sunnypilot.modeld_v2.camera_offset_helper import CameraOffsetHelper
 from openpilot.sunnypilot.livedelay.helpers import get_lat_delay
 from openpilot.sunnypilot.modeld_v2.modeld_base import ModelStateBase
-from openpilot.sunnypilot.modeld_v2.helpers import load_oob
+from openpilot.sunnypilot.modeld_v2.helpers import load_model
+from openpilot.sunnypilot.modeld_v2.frame_fusion import fused_with_warp
 from openpilot.sunnypilot.modeld_v2.model_adapters import get_model_adapter
 from openpilot.sunnypilot.models.helpers import get_active_bundle
 from openpilot.sunnypilot.selfdrive.controls.lib.relc import RoadEdgeLaneChangeController
@@ -52,8 +52,7 @@ BIG_MODEL_TIMEOUT = 60
 
 
 def _pkl_exists(path):
-  from openpilot.common.file_chunker import get_manifest_path
-  return os.path.exists(path) or os.path.exists(get_manifest_path(path))
+  return os.path.exists(path)
 
 
 def _find_driving_pkl(bundle):
@@ -108,7 +107,7 @@ class ModelState(ModelStateBase):
 
   def _init_combined(self, pkl_path, cam_w, cam_h, bundle):
     cloudlog.warning(f"loading combined pkl: {pkl_path}")
-    jits = load_oob(open_file_chunked(pkl_path))
+    jits = load_model(pkl_path, extend=fused_with_warp(cam_w, cam_h, self.chestnut))
 
     metadata = jits['metadata']
     self.WARP_DEV = metadata.get('warp_dev', 'QCOM') if COMMA_HARDWARE else 'CPU'
@@ -258,7 +257,8 @@ def main(demo=False):
 
   CHESTNUT = chestnut_present()
   if CHESTNUT:
-    os.environ['HCQDEV_WAIT_TIMEOUT_MS'] = '3000'
+    from tinygrad.runtime.ops_amd import AMDDevice
+    AMDDevice.wait_timeout_ms = 3000
 
   params = Params()
   params.put_bool("ChestnutLoading", CHESTNUT)

openpilot/sunnypilot/modeld_v2/models/big_driving_warp_1344x760_tinygrad.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ebac006f5e35ddb9c7da8ab38f9af87d6d0a87a3299b07cd118f88581f0f561f
-size 872573
+oid sha256:f1689356ff8e516f87e3d75f6b53e66953a11ee5a97e859b4808c9a16bfe671b
+size 3677806

openpilot/sunnypilot/modeld_v2/models/big_driving_warp_1928x1208_tinygrad.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d8eab79e53ebe1bbc3c633acc2a6bee80612cfe7fb5246afff6e0b7259c3a0d3
-size 872317
+oid sha256:56cc9690e0e9fb719323d2c6646c1305e8bc6bbdfdc807724a605a0d0a37554b
+size 7904366

openpilot/sunnypilot/modeld_v2/models/driving_warp_1344x760_tinygrad.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c48c9150afd01bd7071641415a248308c9a7b3cb47e7095862ef66c9f0429e43
-size 838183
+oid sha256:4b22fd6bcb3a3e8a642e6866449dd9c30c12e694ff0741cad9e34e4b00d8c5fa
+size 3698799

openpilot/sunnypilot/modeld_v2/models/driving_warp_1928x1208_tinygrad.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:59564a7d214fe13dbb30cdd068437c4bf35ec071789c922288046636f711cdfa
-size 837991
+oid sha256:9a17fd6acece4fab23784ef0f3f8bba28a049e9599e5e9e5ee7b36e4b1d24a6f
+size 7924847

openpilot/sunnypilot/modeld_v2/plain_artifact_compiler.py

@@ -0,0 +1,90 @@
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+import dataclasses
+import mmap
+import tempfile
+from contextlib import contextmanager
+from unittest import mock
+
+import numpy as np
+
+from openpilot.sunnypilot.modeld_v2.helpers import Blob, flatten
+
+# Compile time only/ see discountchubbs 4 assistance: turns a captured TinyJit into plain data that helpers.load_model rebuilds. It reads internals of the
+# tinygrad that compiles the model, so a tinygrad bump that breaks it blocks new builds, never loading already built models.
+@contextmanager
+def record_kernel_linear():
+  import tinygrad.engine.jit as jit
+  import tinygrad.engine.realize as realize
+  recorded, hcq_compile, compile_linear = [], realize.hcq_compile, jit.compile_linear
+
+  def recording_compile_linear(*args, **kwargs):
+    with mock.patch.object(realize, "hcq_compile", lambda linear, *a, **k: (recorded.append(linear), hcq_compile(linear, *a, **k))[1]):
+      return compile_linear(*args, **kwargs)
+
+  with mock.patch.object(jit, "compile_linear", recording_compile_linear):
+    yield recorded
+
+
+def capture_frame(jit, kwargs: dict, output_container: str | None = None):
+  inputs = [([name] if key is None else [name, key], tensor.numpy().copy(), tensor.device)
+            for name, value in kwargs.items() for key, tensor in (value.items() if isinstance(value, dict) else [(None, value)])]
+  result = jit(**kwargs)
+  outputs = kwargs[output_container].values() if output_container else flatten(result)
+  return {"inputs": inputs, "outputs": [tensor.numpy().copy() for tensor in outputs], "output_container": output_container}, result
+
+
+def serialize_jit(jit, kernel_linear, check: dict | None = None) -> tuple[dict, memoryview]:
+  from tinygrad.uop.ops import Ops
+  sink = tempfile.TemporaryFile(dir=".")
+  blob, bases, input_specs = Blob(sink), {}, {}
+
+  def store(array) -> dict:
+    array = np.ascontiguousarray(array)
+    return {"data": blob.add(array.tobytes()), "dtype": array.dtype.str, "shape": list(array.shape)}
+
+  def describe(uop) -> dict:
+    if uop.op is Ops.PARAM:
+      input_specs.setdefault(uop.arg.slot, {"dtype": repr(uop.dtype), "numel": int(uop.max_numel()), "device": uop.device})
+      return {"input": uop.arg.slot}
+    buffer = uop.buffer
+    root = buffer._base if buffer._base is not None else buffer
+    base = bases.setdefault(id(root), (len(bases), root, uop.base.dtype if buffer._base is None else uop.dtype))
+    return {"buffer": base[0], "offset": buffer.offset if buffer._base is not None else 0, "nbytes": buffer.nbytes, "dtype": repr(uop.dtype)}
+
+  calls, written, needs_init = [], set(), set()
+  for call in kernel_linear.src:
+    ast, arguments = call.src[0], [describe(argument) for argument in call.src[1:]]
+    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),
+                    "signature": [[name, slot, repr(dtype), [int(x) for x in shape]] for name, slot, dtype, shape in elf.signature],
+                    "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})
+      outs, ins = info.outs, info.ins
+    elif ast.op is Ops.STORE:  # a copy: STORE(PARAM dst, PARAM src)
+      calls.append({"kind": "copy", "args": arguments})
+      outs, ins = (0,), (1,)
+    else:
+      raise ValueError(f"unsupported call {ast.op} {ast.arg}")
+    needs_init |= {arguments[i]["buffer"] for i in ins if "buffer" in arguments[i] and arguments[i]["buffer"] not in written}
+    written |= {arguments[i]["buffer"] for i in outs if "buffer" in arguments[i]}
+
+  ret = jit.captured.ret
+  returns = {"kind": type(ret).__name__, "tensors": [describe(t.uop.base) | {"shape": [int(x) for x in t.shape]} for t in flatten(ret)]}
+  buffers = [{"nbytes": base.nbytes, "device": base.device, "dtype": repr(dtype)}
+             | ({"init": blob.add(base.as_memoryview())} if index in needs_init else {}) for index, base, dtype in bases.values()]
+  names = list(jit.captured.expected_names)
+  if check is not None:
+    check = {"inputs": [[path, store(array), device] for path, array, device in check["inputs"]], "outputs": [store(array) for array in check["outputs"]],
+             "output_container": check["output_container"]}
+  header = {"input_names": names, "buffers": buffers, "calls": calls, "returns": returns, "check": check,
+            "inputs": [{"slot": slot, "name": names[slot] if slot < len(names) else None} | input_specs[slot] for slot in sorted(input_specs)]}
+  with sink:
+    sink.flush()
+    return header, memoryview(mmap.mmap(sink.fileno(), 0, access=mmap.ACCESS_READ) if blob.size else b"")

openpilot/sunnypilot/modeld_v2/tests/helpers.py

@@ -188,10 +188,8 @@ def make_pkl_data(archetype):
 
 
 def write_pkl(tmp_path, archetype):
-  from openpilot.sunnypilot.modeld_v2.helpers import dump_oob
   pkl_path = tmp_path / 'driving_test_tinygrad.pkl'
-  with open(pkl_path, 'wb') as f:
-    dump_oob(make_pkl_data(archetype), f)
+  pkl_path.write_bytes(pickle.dumps(make_pkl_data(archetype)))
   return pkl_path
 
 
@@ -211,6 +209,7 @@ def patch_modeld(monkeypatch):
   def _patch(bundle):
     monkeypatch.setattr(helpers, 'get_active_bundle', lambda params=None, *, chestnut=None: bundle)
     monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None, *, chestnut=None: bundle)
+    monkeypatch.setattr(modeld_module, 'load_model', lambda path, **_: pickle.loads(pathlib.Path(path).read_bytes()))
 
   return _patch
 

openpilot/sunnypilot/modeld_v2/tests/test_combined_pkl_loader.py

@@ -90,7 +90,6 @@ class TestStockEquivalence(OpenpilotTestCase):
 
   def test_unified_run_model(self, tmp_path, monkeypatch, patch_modeld):
     from openpilot.common.hardware import hw
-    from openpilot.sunnypilot.modeld_v2.helpers import dump_oob
     shapes = {'img': (1, 12, 128, 256), 'big_img': (1, 12, 128, 256), 'features_buffer': (1, 24, 32, 512),
               'desire_pulse': (1, 25, 8), 'traffic_convention': (1, 2), 'action_t': (1, 2)}
     import codecs
@@ -102,8 +101,7 @@ class TestStockEquivalence(OpenpilotTestCase):
       (CAM_W, CAM_H): tests_helpers._noop_jit,
       'run_policy': tests_helpers._noop_jit,
     }
-    with open(tmp_path / 'driving_test_tinygrad.pkl', 'wb') as f:
-      dump_oob(pkl_data, f)
+    (tmp_path / 'driving_test_tinygrad.pkl').write_bytes(pickle.dumps(pkl_data))
     bundle = DummyBundle(models=[DummyModel('supercombo', 'driving_test_tinygrad.pkl')])
     patch_modeld(bundle)
     monkeypatch.setattr(hw.Paths, 'model_root', staticmethod(lambda: str(tmp_path)))

openpilot/sunnypilot/modeld_v2/tests/test_compile_modeld.py

@@ -5,16 +5,12 @@ This file is part of sunnypilot and is licensed under the MIT License.
 See the LICENSE.md file in the root directory for more details.
 """
 
-import os
-import tempfile
 import unittest
-from pathlib import Path
 
 import numpy as np
 from openpilot.common.parameterized import parameterized
 
-from openpilot.common.file_chunker import chunk_file, get_chunk_targets
-from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, _detect_desire_key, read_file_chunked_to_disk
+from openpilot.sunnypilot.modeld_v2.compile_modeld import derive_frame_skip, _detect_desire_key
 from openpilot.common.test import OpenpilotTestCase
 
 
@@ -168,36 +164,6 @@ class TestOutputSlicePreservation(OpenpilotTestCase):
       "vision and policy slices should not overlap in keys"
 
 
-class TestReadFileChunkedToDisk(OpenpilotTestCase):
-  def test_none_passthrough(self):
-    assert read_file_chunked_to_disk(None) is None
-
-  def test_unchunked_source_staged_on_disk(self):
-    with tempfile.TemporaryDirectory() as d:
-      src = Path(d) / "driving_supercombo.onnx"
-      payload = os.urandom(1024)
-      src.write_bytes(payload)
-
-      out = Path(read_file_chunked_to_disk(str(src)))
-
-      assert out.parent == Path(d)
-      assert out.name == "driving_supercombo.onnx.unchunked"
-      assert out.read_bytes() == payload
-
-  def test_chunked_source_reassembled_on_disk(self):
-    with tempfile.TemporaryDirectory() as d:
-      src = Path(d) / "driving_supercombo.onnx"
-      payload = os.urandom(4096)
-      src.write_bytes(payload)
-      chunk_file(str(src), get_chunk_targets(str(src), len(payload)))
-      assert not src.exists()
-
-      out = Path(read_file_chunked_to_disk(str(src)))
-
-      assert out.parent == Path(d)
-      assert out.read_bytes() == payload
-
-
 class Test4DFeaturesBuffer(OpenpilotTestCase):
   def test_get_policy_npy_shapes_4d(self):
     from openpilot.sunnypilot.modeld_v2.compile_modeld import get_policy_npy_shapes

openpilot/sunnypilot/modeld_v2/tests/test_fallback.py

@@ -5,13 +5,11 @@ This file is part of sunnypilot and is licensed under the MIT License.
 See the LICENSE.md file in the root directory for more details.
 """
 
-import io
+import pickle
 import requests
 
-from openpilot.common.file_chunker import get_chunk_name
 from openpilot.common.hardware import hw
 from openpilot.common.test import OpenpilotTestCase
-from openpilot.sunnypilot.modeld_v2.helpers import dump_oob
 import openpilot.sunnypilot.modeld_v2.modeld as modeld_module
 from openpilot.sunnypilot.modeld_v2.tests import helpers as tests_helpers
 from openpilot.sunnypilot.modeld_v2.tests.helpers import DummyModel, DummyBundle, CAM_W, CAM_H
@@ -43,14 +41,11 @@ class TestFallback(OpenpilotTestCase):
     small_json = requests.get(ModelFetcher.MODEL_URL).json()
     small_bundle = ModelParser.parse_models(small_json)[-1]
 
-    buf = io.BytesIO()
-    dump_oob(tests_helpers.make_pkl_data(tests_helpers.ARCHETYPES['supercombo_non20hz']), buf)
-    oob_bytes = buf.getvalue()
+    oob_bytes = pickle.dumps(tests_helpers.make_pkl_data(tests_helpers.ARCHETYPES['supercombo_non20hz']))
+    monkeypatch.setattr(modeld_module, 'load_model', lambda path, **_: pickle.loads(open(path, 'rb').read()))
 
     for bundle in (big_bundle, small_bundle):
-      artifact = bundle.models[0].artifact
-      for i in range(len(artifact.chunks)):
-        (tmp_path / get_chunk_name(artifact.fileName, i, len(artifact.chunks))).write_bytes(oob_bytes if i == 0 else b"")
+      (tmp_path / bundle.models[0].artifact.fileName).write_bytes(oob_bytes)
 
     monkeypatch.setattr(modeld_module, 'get_active_bundle', lambda params=None, *, chestnut=None: small_bundle)
     assert modeld_module.ModelState(CAM_W, CAM_H, chestnut=False).chestnut is False

openpilot/sunnypilot/modeld_v2/tests/test_models.py

@@ -7,71 +7,14 @@ See the LICENSE.md file in the root directory for more details.
 
 import os
 import unittest
-from unittest.mock import patch
-from openpilot.common.file_chunker import open_file_chunked
-from openpilot.sunnypilot.modeld_v2.helpers import load_oob
-from tinygrad.device import Device
+
+from openpilot.sunnypilot.modeld_v2.helpers import PlainJit, load_model
 
 
-class TestLegacyModels(unittest.TestCase):
-  def test_legacy_model_load(self):
-    base_name = os.environ.get("MODEL_BASE_NAME")
-    if not base_name:
-      raise unittest.SkipTest("MODEL_BASE_NAME env var not set, skipping integration test.")
-    chunk_dir = os.environ.get("MODEL_CHUNK_DIR", "/tmp/model_chunks")
-    base_path = os.path.join(chunk_dir, base_name)
-
-    try:
-      f = open_file_chunked(base_path)
-    except Exception as error:
-      self.fail(f"Failed to open chunked file {base_path}: {error}")
-    self.addCleanup(f.close)
-
-    real_getitem = Device.__class__.__getitem__
-
-    def safe_getitem(device_self, ix):
-      if ix == "QCOM" and not os.path.exists("/dev/kgsl-3d0"):
-        return real_getitem(device_self, "CPU")
-      if ix == "AMD" and not os.path.exists("/dev/kfd"):
-        return real_getitem(device_self, "CPU")
-      return real_getitem(device_self, ix)
-
-    with patch.object(Device.__class__, "__getitem__", safe_getitem):
-      obj = load_oob(f)
-
-      assert isinstance(obj, dict), "Parsed object is not a dictionary"
-      assert "metadata" in obj, "Metadata key is missing"
-
-      def traverse_jits(d):
-        if hasattr(d, "captured") and d.captured is not None:
-          linear_attr = getattr(d.captured, "_linear", None)
-          if linear_attr is None:
-            try:
-              linear_attr = getattr(d.captured, "linear", None)
-            except AssertionError as e:
-              if "alloc size must be positive" in str(e):
-                return
-              raise
-
-          if linear_attr is not None:
-            stack = [linear_attr]
-            visited = set()
-            while stack:
-              node = stack.pop()
-              if id(node) in visited:
-                continue
-              visited.add(id(node))
-              if hasattr(node, "src") and node.src is not None:
-                for child in node.src:
-                  stack.append(child)
-        elif isinstance(d, dict):
-          for v in d.values():
-            traverse_jits(v)
-        elif isinstance(d, (list, tuple)):
-          for v in d:
-            traverse_jits(v)
-
-      if "run_model" in obj:
-        traverse_jits(obj["run_model"])
-      if "run_policy" in obj:
-        traverse_jits(obj["run_policy"])
+class TestRegistryModels(unittest.TestCase):
+  def test_model_loads(self):
+    if not (base_name := os.environ.get("MODEL_BASE_NAME")):
+      self.skipTest("MODEL_BASE_NAME env var not set, skipping integration test.")
+    bundle = load_model(os.path.join(os.environ.get("MODEL_DIR", "/tmp/models"), base_name), {"QCOM": "CPU", "AMD": "CPU"}, verify=False, cache=False)
+    jits = [value for value in bundle.values() if isinstance(value, PlainJit)]
+    assert "metadata" in bundle and jits and all(len(jit.captured.linear.src) > 0 for jit in jits)

openpilot/sunnypilot/modeld_v2/warmup.py

@@ -0,0 +1,42 @@
+#!/usr/bin/env python3
+"""
+Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
+
+This file is part of sunnypilot and is licensed under the MIT License.
+See the LICENSE.md file in the root directory for more details.
+"""
+
+import os
+os.environ['GMMU'] = '0'
+import multiprocessing
+import time
+from openpilot.common.hardware.hw import Paths
+from openpilot.common.params import Params
+from openpilot.common.swaglog import cloudlog
+from openpilot.sunnypilot.modeld_v2.helpers import cache_is_fresh, load_model, read_parts
+from openpilot.sunnypilot.models.helpers import get_selected_bundle
+
+
+def main():
+  """
+  The main entry point for offroad cache creation.
+  Runs every 10 seconds and checks for a stale qcom model cache.
+  """
+  params, failed, native, spawn = Params(), set(), set(), multiprocessing.get_context("spawn")
+  while True:
+    bundle = get_selected_bundle(params, "qcom")
+    path = os.path.join(Paths.model_root(), bundle.models[0].artifact.fileName) if bundle and bundle.models else ""
+    if os.path.isfile(path) and not cache_is_fresh(path) and (attempt := (path, os.stat(path).st_mtime_ns)) not in failed | native:
+      if "input_specs" in read_parts(path)[1]:  # native models are fused
+        native.add(attempt)
+      else:
+        (worker := spawn.Process(target=load_model, args=(path,), daemon=True)).start()
+        worker.join()
+        if not cache_is_fresh(path):
+          cloudlog.error(f"model cache warmup failed for {path} with exit code {worker.exitcode}")
+          failed.add(attempt)
+    time.sleep(10)
+
+
+if __name__ == "__main__":
+  main()

openpilot/sunnypilot/models/fetcher.py

@@ -6,12 +6,10 @@ See the LICENSE.md file in the root directory for more details.
 """
 
 import time
-import os
 import requests
 from requests.exceptions import (SSLError, RequestException, HTTPError)
 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.cereal import custom
 
@@ -26,35 +24,11 @@ class ModelParser:
     download_uri.sha256 = download_uri_data.get("sha256")
     return download_uri
 
-  @staticmethod
-  def _parse_chunk(chunk_data) -> custom.ModelManagerSP.Chunk:
-    chunk = custom.ModelManagerSP.Chunk()
-    chunk.fileName = chunk_data.get("file_name")
-    chunk.sha256 = chunk_data.get("sha256")
-    return chunk
-
   @staticmethod
   def _parse_artifact(artifact_data) -> custom.ModelManagerSP.Artifact:
     artifact = custom.ModelManagerSP.Artifact()
     artifact.fileName = artifact_data.get("file_name")
     artifact.downloadUri = ModelParser._parse_download_uri(artifact_data.get("download_uri", {}))
-
-    if "chunks" in artifact_data:
-      artifact.chunks = [ModelParser._parse_chunk(chunk_data) for chunk_data in artifact_data["chunks"]]
-
-      try:
-        model_dir = Paths.model_root()
-        os.makedirs(model_dir, exist_ok=True)
-        manifest_path = os.path.join(model_dir, f"{artifact.fileName}.chunkmanifest")
-        num_chunks = str(len(artifact.chunks))
-
-        if not os.path.exists(manifest_path) or open(manifest_path).read().strip() != num_chunks:
-          with open(manifest_path, "w") as f:
-            f.write(num_chunks)
-          cloudlog.info(f"Wrote chunk manifest for {artifact.fileName}: {num_chunks} chunks")
-      except Exception as e:
-        cloudlog.warning(f"Failed to write chunk manifest for {artifact.fileName}: {e}")
-
     return artifact
 
   @staticmethod
@@ -138,8 +112,8 @@ class ModelCache:
 
 class ModelFetcher:
   """Handles fetching and caching of model data from remote source"""
-  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v23.json"
-  MODEL_URL_CHESTNUT = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_chestnut_v27.json"
+  MODEL_URL = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_v25.json"
+  MODEL_URL_CHESTNUT = "https://raw.githubusercontent.com/sunnypilot/sunnypilot-models/refs/heads/gh-pages/docs/driving_models_chestnut_v28.json"
 
   MODEL_SOURCES = {
     "qcom": (MODEL_URL, ""),
@@ -270,5 +244,3 @@ if __name__ == "__main__":
       model_overrides = {override.key: override.value for override in bundle.overrides}
       print(f"Bundle: {bundle.internalName}, Type: {model.type}, Status: {bundle.status}, Overrides: {model_overrides}")
       print(f"Artifact: {model.artifact.fileName}, Download URI: {model.artifact.downloadUri.uri}")
-      if model.artifact.chunks:
-        print(f"Contains {len(model.artifact.chunks)} chunks.")

openpilot/sunnypilot/models/helpers.py

@@ -16,7 +16,7 @@ from openpilot.common.hardware.hw import Paths
 from openpilot.selfdrive.modeld.helpers import chestnut_present
 
 # SET ME TO THE EXACT JSON VERSION WE SET IN SUNNYPILOT_MODELS REPO
-REQUIRED_JSON_VERSION = 20
+REQUIRED_JSON_VERSION = 21
 
 CUSTOM_MODEL_PATH = Paths.model_root()
 ModelManager = custom.ModelManagerSP
@@ -29,9 +29,8 @@ _LAST_VALIDATED_RAW: dict[str, dict | None] = {}
 
 
 def _compute_hash(file_path: str) -> str | None:
-  from openpilot.common.file_chunker import open_file_chunked
   try:
-    with open_file_chunked(file_path) as file:
+    with open(file_path, 'rb') as file:
       return hashlib.file_digest(file, "sha256").hexdigest().lower()
   except FileNotFoundError:
     return None
@@ -58,20 +57,11 @@ def is_bundle_version_compatible(bundle: dict) -> bool:
 
 def _bundle_artifacts(bundle: custom.ModelManagerSP.ModelBundle) -> list[tuple[str, str]]:
   artifacts = []
-  from openpilot.common.file_chunker import get_chunk_name
   for model in getattr(bundle, 'models', []) or []:
-    for artifact in (getattr(model, 'artifact', None),):
-      if artifact and getattr(artifact, 'fileName', None):
-        if len(artifact.chunks) > 0:
-          for i, chunk in enumerate(artifact.chunks):
-            chunk_name = get_chunk_name(artifact.fileName, i, len(artifact.chunks))
-            if getattr(chunk, 'sha256', None):
-              artifacts.append((chunk_name, chunk.sha256))
-        else:
-          if getattr(artifact, 'downloadUri', None):
-            sha256 = getattr(artifact.downloadUri, 'sha256', None)
-            if sha256:
-              artifacts.append((artifact.fileName, sha256))
+    artifact = getattr(model, 'artifact', None)
+    if artifact and getattr(artifact, 'fileName', None) and getattr(artifact, 'downloadUri', None):
+      if sha256 := getattr(artifact.downloadUri, 'sha256', None):
+        artifacts.append((artifact.fileName, sha256))
   return artifacts
 
 

openpilot/sunnypilot/models/manager.py

@@ -109,50 +109,6 @@ class ModelManagerSP:
     # Clean up start time after download completes
     del self._download_start_times[model.fileName]
 
-  async def _download_chunked(self, base_url: str, base_path: str, artifact, skip: frozenset[int] | set[int] = frozenset()) -> None:
-    from openpilot.common.file_chunker import get_chunk_name, get_manifest_path
-
-    num_chunks = len(artifact.chunks)
-    if num_chunks == 0:
-      raise ValueError("No chunks defined in artifact")
-
-    manifest_path = get_manifest_path(base_path)
-    self._download_start_times[artifact.fileName] = time.monotonic()
-
-    # Shared connection saves a TCP+TLS handshake per chunk.
-    # Keep sequential: the link saturates on one stream and Session is not thread-safe.
-    completed = len(skip)
-    with requests.Session() as session:
-      for i, _ in enumerate(artifact.chunks):
-        if i in skip:
-          continue
-        chunk_url = get_chunk_name(base_url, i, num_chunks)
-        chunk_path = get_chunk_name(base_path, i, num_chunks)
-        chunk_downloaded = 0
-        with session.get(chunk_url, stream=True, timeout=DOWNLOAD_TIMEOUT) as response:
-          response.raise_for_status()
-          chunk_size = int(response.headers.get("content-length", 0))
-          with open(chunk_path, 'wb') as f:  # noqa: ASYNC230
-            for data in response.iter_content(chunk_size=self._chunk_size):
-              f.write(data)
-              chunk_downloaded += len(data)
-              if self._download_interrupted():
-                raise DownloadCancelled("Download cancelled")
-              intra = chunk_downloaded / max(chunk_size, 1)
-              progress = min(99.0, ((completed + intra) / num_chunks) * 100)
-              artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloading
-              artifact.downloadProgress.progress = progress
-              artifact.downloadProgress.eta = self._calculate_eta(artifact.fileName, progress)
-              self._sync_artifact_progress(artifact)
-              self._report_status()
-        completed += 1
-
-    with open(manifest_path, 'w') as f:  # noqa: ASYNC230
-      f.write(str(num_chunks))
-    if os.path.isfile(base_path):  # noqa: ASYNC240
-      os.remove(base_path)
-    del self._download_start_times[artifact.fileName]
-
   async def _process_artifact(self, artifact, destination_path: str) -> None:
     if not artifact.downloadUri.uri:
       return None
@@ -165,28 +121,7 @@ class ModelManagerSP:
     full_path = os.path.join(destination_path, filename)
 
     try:
-      # progress counts only valid chunks so a resumed download continues the
-      # bar from where verification left it, instead of falling back to zero
-      is_cached = False
-      valid_chunks: set[int] = set()
-      if len(artifact.chunks) > 0:
-        from openpilot.common.file_chunker import get_chunk_name
-        num_chunks = len(artifact.chunks)
-        for i, chunk in enumerate(artifact.chunks):
-          if self._download_interrupted():
-            raise DownloadCancelled("Download cancelled")
-          if await verify_file(get_chunk_name(full_path, i, num_chunks), chunk.sha256):
-            valid_chunks.add(i)
-          artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.verifying
-          artifact.downloadProgress.progress = (len(valid_chunks) / num_chunks) * 100
-          self._sync_artifact_progress(artifact)
-          self._report_status()
-        is_cached = len(valid_chunks) == num_chunks
-      else:
-        if await verify_file(full_path, expected_hash):
-          is_cached = True
-
-      if is_cached:
+      if await verify_file(full_path, expected_hash):
         artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.cached
         artifact.downloadProgress.progress = 100
         artifact.downloadProgress.eta = 0
@@ -194,17 +129,9 @@ class ModelManagerSP:
         self._report_status()
         return
 
-      if len(artifact.chunks) > 0:
-        await self._download_chunked(url, full_path, artifact, skip=valid_chunks)
-        from openpilot.common.file_chunker import get_chunk_name
-        for i, chunk in enumerate(artifact.chunks):
-          chunk_path = get_chunk_name(full_path, i, len(artifact.chunks))
-          if not await verify_file(chunk_path, chunk.sha256):
-            raise ValueError(f"Hash validation failed for chunk {i+1} of {filename}")
-      else:
-        await self._download_file(url, full_path, artifact)
-        if not await verify_file(full_path, expected_hash):
-          raise ValueError(f"Hash validation failed for {filename}")
+      await self._download_file(url, full_path, artifact)
+      if not await verify_file(full_path, expected_hash):
+        raise ValueError(f"Hash validation failed for {filename}")
 
       artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.downloaded
       artifact.downloadProgress.progress = 100
@@ -213,7 +140,6 @@ class ModelManagerSP:
       self._report_status()
 
     except DownloadCancelled:
-      # a cancel keeps whatever is on disk: complete chunks resume the next attempt
       self._download_start_times.pop(artifact.fileName, None)
       artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
       artifact.downloadProgress.eta = 0
@@ -225,9 +151,8 @@ class ModelManagerSP:
 
     except Exception as e:
       cloudlog.error(f"Error downloading {filename}: {str(e)}")
-      for f in [full_path] + [p for p in (os.path.join(destination_path, f) for f in os.listdir(destination_path)) if filename in p]:
-        if os.path.isfile(f):  # noqa: ASYNC240
-          os.remove(f)
+      if os.path.isfile(full_path):  # noqa: ASYNC240
+        os.remove(full_path)
       artifact.downloadProgress.status = custom.ModelManagerSP.DownloadStatus.failed
       artifact.downloadProgress.eta = 0
       self._sync_artifact_progress(artifact)
@@ -360,12 +285,11 @@ class ModelManagerSP:
           if model.artifact.fileName:
             active_files.append(model.artifact.fileName)
 
-    # Remove all files except active ones (including their chunk files)
+    # Remove all files except active ones
     model_dir = Paths.model_root()
     try:
       for filename in os.listdir(model_dir):
-        base = filename.split('.chunk')[0] if '.chunk' in filename else filename
-        if base not in active_files and filename not in active_files:
+        if filename not in active_files:
           file_path = os.path.join(model_dir, filename)
           if os.path.isfile(file_path):
             os.remove(file_path)

openpilot/sunnypilot/models/tests/test_manager_download.py

@@ -17,11 +17,9 @@ from typing import Any
 from unittest import mock
 
 import requests
-from urllib3.connectionpool import HTTPConnectionPool
 
 from openpilot.cereal import custom
 from openpilot.common.test import OpenpilotTestCase
-from openpilot.common.file_chunker import get_chunk_name, get_manifest_path
 from openpilot.selfdrive.test.helpers import http_server_context
 from openpilot.sunnypilot.models import manager as manager_module
 from openpilot.sunnypilot.models.fetcher import ModelFetcher, get_cached_bundles
@@ -30,7 +28,6 @@ from openpilot.sunnypilot.models.helpers import (get_active_bundle, get_active_s
                                                   resolve_bundle_by_ref, validate_active_bundles)
 from openpilot.sunnypilot.models.manager import ModelManagerSP
 
-CHUNK_BODIES = [b'A' * 5000, b'B' * 5000, b'C' * 3000]
 WHOLE_BODY = b'Z' * 9000
 
 
@@ -49,12 +46,7 @@ class DownloadHandler(http.server.BaseHTTPRequestHandler):
     pass
 
   def _body_for(self, path):
-    if path.endswith('.whole'):
-      return WHOLE_BODY
-    for i in range(len(CHUNK_BODIES)):
-      if path.endswith(get_chunk_name('', i, len(CHUNK_BODIES))):
-        return CHUNK_BODIES[i]
-    return None
+    return WHOLE_BODY if path.endswith('.pkl') else None
 
   def do_GET(self):
     type(self).request_paths.append(self.path)
@@ -119,35 +111,19 @@ class ManagerDownloadTestBase(OpenpilotTestCase):
     if artifact is not None:
       self.reported.append(float(artifact.downloadProgress.progress))
 
-  def make_artifact(self, chunked: bool):
+  def make_artifact(self):
     bundle = custom.ModelManagerSP.ModelBundle.new_message()
     bundle.init('models', 1)
     artifact = bundle.models[0].artifact
     artifact.fileName = 'driving_test_tinygrad.pkl'
-    if chunked:
-      artifact.downloadUri.uri = self.base_url + '/driving_test_tinygrad.pkl'
-      artifact.downloadUri.sha256 = sha256(b''.join(CHUNK_BODIES))
-      artifact.init('chunks', len(CHUNK_BODIES))
-      for i, body in enumerate(CHUNK_BODIES):
-        artifact.chunks[i].sha256 = sha256(body)
-    else:
-      artifact.downloadUri.uri = self.base_url + '/driving_test_tinygrad.pkl.whole'
-      artifact.downloadUri.sha256 = sha256(WHOLE_BODY)
+    artifact.downloadUri.uri = self.base_url + '/driving_test_tinygrad.pkl'
+    artifact.downloadUri.sha256 = sha256(WHOLE_BODY)
     self._bundle = bundle
     self.artifact = artifact
     return artifact
 
-  def chunk_paths(self, base_path):
-    return [get_chunk_name(base_path, i, len(CHUNK_BODIES)) for i in range(len(CHUNK_BODIES))]
-
-  def assert_no_partials(self, base_path):
-    leftovers = [p for p in [base_path, get_manifest_path(base_path)] + self.chunk_paths(base_path)
-                 if os.path.isfile(p)]
-    assert leftovers == [], f"partial files left behind: {leftovers}"
-
-
 class TestManagerDownload(ManagerDownloadTestBase):
-  """Exercises the real _download_file / _download_chunked against a local server."""
+  """Exercises the real _download_file against a local server."""
 
   def run_with_server(self, fn):
     with http_server_context(handler=DownloadHandler) as (host, port):
@@ -156,7 +132,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
 
   def test_download_file_writes_exact_bytes(self):
     def body():
-      artifact = self.make_artifact(chunked=False)
+      artifact = self.make_artifact()
       path = os.path.join(self.dest, artifact.fileName)
       asyncio.run(self.manager._download_file(artifact.downloadUri.uri, path, artifact))
       with open(path, 'rb') as f:
@@ -166,111 +142,67 @@ class TestManagerDownload(ManagerDownloadTestBase):
       assert artifact.fileName not in self.manager._download_start_times
     self.run_with_server(body)
 
-  def test_download_chunked_writes_all_chunks_and_manifest(self):
-    def body():
-      artifact = self.make_artifact(chunked=True)
-      base_path = os.path.join(self.dest, artifact.fileName)
-      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-
-      for i, expected in enumerate(CHUNK_BODIES):
-        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'rb') as f:
-          assert f.read() == expected, f"chunk {i} body mismatch"
-
-      with open(get_manifest_path(base_path)) as f:
-        assert f.read() == str(len(CHUNK_BODIES))
-
-      assert not os.path.isfile(base_path), "base file should be removed after chunking"
-      assert artifact.fileName not in self.manager._download_start_times
-    self.run_with_server(body)
-
   def test_progress_is_monotonic_and_bounded(self):
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
-      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
+      asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
 
       assert self.reported, "expected progress reports"
       for a, b in zip(self.reported, self.reported[1:], strict=False):
         assert b >= a, f"progress went backwards: {a} -> {b}"
-      assert max(self.reported) <= 99.0, f"chunked progress must stay <=99 until verify, got {max(self.reported)}"
-    self.run_with_server(body)
-
-  def test_session_is_reused_across_chunks(self):
-    """One connection pool shared across every chunk."""
-    def body():
-      artifact = self.make_artifact(chunked=True)
-      base_path = os.path.join(self.dest, artifact.fileName)
-
-      pools = []
-      original = HTTPConnectionPool.urlopen
-
-      def tracked(pool_self, *args, **kwargs):
-        pools.append(id(pool_self))
-        return original(pool_self, *args, **kwargs)
-
-      with mock.patch.object(HTTPConnectionPool, 'urlopen', tracked):
-        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-
-      assert len(pools) == len(CHUNK_BODIES), f"expected one request per chunk, got {len(pools)}"
-      assert len(set(pools)) == 1, f"connection pool not reused across chunks: {len(set(pools))} pools"
     self.run_with_server(body)
 
   def test_http_error_propagates(self):
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
-      failing = '/' + os.path.basename(get_chunk_name(artifact.downloadUri.uri, 1, len(CHUNK_BODIES)))
-      DownloadHandler.fail_paths = {failing: 404}
+      DownloadHandler.fail_paths = {'/' + artifact.fileName: 404}
 
       with self.assertRaises(requests.exceptions.HTTPError):
-        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-
-      # chunk 1 failed, so its file and the manifest must not exist
-      assert not os.path.isfile(get_chunk_name(base_path, 1, len(CHUNK_BODIES)))
-      assert not os.path.isfile(get_manifest_path(base_path))
+        asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
+      assert not os.path.isfile(base_path)
     self.run_with_server(body)
 
   def test_cancellation_mid_transfer(self):
     """Cancellation is checked inside the byte loop; it must still fire after the port."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
       self.manager.params.get.return_value = None  # cancelled
 
       with self.assertRaises(Exception) as ctx:
-        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
+        asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
       assert 'cancelled' in str(ctx.exception).lower()
-      assert not os.path.isfile(get_manifest_path(base_path))
     self.run_with_server(body)
 
   def test_repeat_downloads_are_stable(self):
     """Back-to-back runs must produce identical bytes and leak no start-time state."""
     def body():
       for _ in range(2):
-        artifact = self.make_artifact(chunked=True)
+        artifact = self.make_artifact()
         base_path = os.path.join(self.dest, artifact.fileName)
-        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-        for i, expected in enumerate(CHUNK_BODIES):
-          with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'rb') as f:
-            assert f.read() == expected
+        asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
+        with open(base_path, 'rb') as f:
+          assert f.read() == WHOLE_BODY
         assert self.manager._download_start_times == {}
     self.run_with_server(body)
 
   def test_download_ref_present_keeps_download_alive(self):
     """A pending download request (DownloadRef set) must not be cancelled mid-transfer."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
       self.manager.params.get.side_effect = lambda key: b"ref" if key == "ModelManager_DownloadRef" else None
       self.manager._download_ref = b"ref"
-      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-      assert os.path.isfile(get_manifest_path(base_path))
+      asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
+      assert os.path.isfile(base_path)
     self.run_with_server(body)
 
   def test_cancellation_via_download_ref(self):
     """Removing DownloadRef mid-transfer cancels the download."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
       checks = {"n": 0}
 
@@ -283,20 +215,19 @@ class TestManagerDownload(ManagerDownloadTestBase):
       self.manager.params.get.side_effect = get
       self.manager._download_ref = b"ref"
       with self.assertRaises(Exception) as ctx:
-        asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
+        asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
       assert 'cancelled' in str(ctx.exception).lower()
-      assert not os.path.isfile(get_manifest_path(base_path))
     self.run_with_server(body)
 
   def test_replaced_download_ref_queues_instead_of_cancelling(self):
     """Selecting another model mid-transfer lets the running download finish."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
       self.manager.params.get.side_effect = lambda key: b"other-ref" if key == "ModelManager_DownloadRef" else None
       self.manager._download_ref = b"ref"
-      asyncio.run(self.manager._download_chunked(artifact.downloadUri.uri, base_path, artifact))
-      assert os.path.isfile(get_manifest_path(base_path))
+      asyncio.run(self.manager._download_file(artifact.downloadUri.uri, base_path, artifact))
+      assert os.path.isfile(base_path)
     self.run_with_server(body)
 
   def test_replaced_download_ref_is_kept(self):
@@ -315,11 +246,10 @@ class TestManagerDownload(ManagerDownloadTestBase):
   def test_cached_bundle_cancel_skips_slot_write(self):
     """A cancel must stop an already-on-disk bundle before it is applied to the slot."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       base_path = os.path.join(self.dest, artifact.fileName)
-      for i, data in enumerate(CHUNK_BODIES):
-        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'wb') as f:
-          f.write(data)
+      with open(base_path, 'wb') as f:
+        f.write(WHOLE_BODY)
       self._bundle.ref = "test-ref"
       params, store = self._make_params_with_store()
       store["ModelManager_DownloadRef"] = None  # removed -> cancelled
@@ -329,42 +259,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
         asyncio.run(self.manager._download_bundle(self._bundle, self.dest, "qcom"))
       assert 'cancelled' in str(ctx.exception).lower()
       assert "ModelManager_ActiveBundle" not in store
-      assert all(os.path.isfile(p) for p in self.chunk_paths(base_path)), "cancel must not delete cached chunks"
-    self.run_with_server(body)
-
-  def test_resume_skips_valid_chunks(self):
-    """A chunk already on disk is kept and not re-downloaded; progress starts above its share."""
-    def body():
-      artifact = self.make_artifact(chunked=True)
-      base_path = os.path.join(self.dest, artifact.fileName)
-      with open(get_chunk_name(base_path, 0, len(CHUNK_BODIES)), 'wb') as f:
-        f.write(CHUNK_BODIES[0])
-
-      asyncio.run(self.manager._process_artifact(artifact, self.dest))
-
-      chunk0_suffix = get_chunk_name('', 0, len(CHUNK_BODIES))
-      assert not any(p.endswith(chunk0_suffix) for p in DownloadHandler.request_paths), "valid chunk was re-downloaded"
-      for i, expected in enumerate(CHUNK_BODIES):
-        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'rb') as f:
-          assert f.read() == expected
-      assert os.path.isfile(get_manifest_path(base_path))
-      assert min(self.reported) >= (1 / len(CHUNK_BODIES)) * 100 - 1, "progress must not restart below the resumed share"
-    self.run_with_server(body)
-
-  def test_verify_reports_valid_fraction_then_cached(self):
-    """A fully cached bundle publishes climbing verify progress and ends cached."""
-    def body():
-      artifact = self.make_artifact(chunked=True)
-      base_path = os.path.join(self.dest, artifact.fileName)
-      for i, data in enumerate(CHUNK_BODIES):
-        with open(get_chunk_name(base_path, i, len(CHUNK_BODIES)), 'wb') as f:
-          f.write(data)
-
-      asyncio.run(self.manager._process_artifact(artifact, self.dest))
-
-      assert DownloadHandler.request_paths == [], "cached bundle must not hit the network"
-      assert [round(p) for p in self.reported[:3]] == [33, 67, 100]
-      assert artifact.downloadProgress.status == custom.ModelManagerSP.DownloadStatus.cached
+      assert os.path.isfile(base_path), "cancel must not delete the cached model"
     self.run_with_server(body)
 
   def _make_params_with_store(self):
@@ -384,7 +279,7 @@ class TestManagerDownload(ManagerDownloadTestBase):
   def test_download_writes_qcom_slot(self):
     """A download resolved to the qcom source writes the qcom active bundle slot only."""
     def body():
-      artifact = self.make_artifact(chunked=True)
+      artifact = self.make_artifact()
       self._bundle.ref = "test-ref"
       self._bundle.minimumSelectorVersion = helpers.REQUIRED_JSON_VERSION
       params, store = self._make_params_with_store()
@@ -396,15 +291,13 @@ class TestManagerDownload(ManagerDownloadTestBase):
       assert self.manager.selected_bundle.status == custom.ModelManagerSP.DownloadStatus.downloaded
       assert self.manager.active_bundle is not None and self.manager.active_bundle.ref == "test-ref"
       assert self.manager.active_bundle.status == custom.ModelManagerSP.DownloadStatus.downloaded
-      chunk_names = [get_chunk_name(artifact.fileName, i, len(artifact.chunks)) for i in range(len(artifact.chunks))]
-      missing = [c for c in chunk_names if not os.path.isfile(os.path.join(self.dest, c))]
-      assert missing == [], f"chunks missing from the cache: {missing}"
+      assert os.path.isfile(os.path.join(self.dest, artifact.fileName))
     self.run_with_server(body)
 
   def test_download_writes_chestnut_slot(self):
     """A download resolved to the chestnut source writes the chestnut active bundle slot only."""
     def body():
-      self.make_artifact(chunked=True)
+      self.make_artifact()
       self._bundle.ref = "big-ref"
       self._bundle.minimumSelectorVersion = helpers.REQUIRED_JSON_VERSION
       params, store = self._make_params_with_store()
@@ -782,7 +675,7 @@ class TestEffectiveSource(OpenpilotTestCase):
 
 @unittest.skipUnless(os.environ.get('RUN_INTEGRATION_TESTS'), 'requires external network')
 class TestLiveModelManifest(OpenpilotTestCase):
-  """Every artifact and chunk URL in the published manifest must resolve."""
+  """Every artifact URL in the published manifest must resolve."""
 
   def test_all_manifest_urls_available(self):
     from openpilot.sunnypilot.models.fetcher import ModelFetcher
@@ -794,10 +687,7 @@ class TestLiveModelManifest(OpenpilotTestCase):
     for bundle in manifest.get('bundles', []):
       for model in bundle.get('models', []):
         artifact = model['artifact']
-        url = artifact['download_uri']['url']
-        chunks = artifact.get('chunks', [])
-        urls = ([url] if not chunks
-                else [get_chunk_name(url, i, len(chunks)) for i in range(len(chunks))])
+        urls = [artifact['download_uri']['url']]
         for u in urls:
           try:
             r = session.head(u, timeout=15, allow_redirects=True)

openpilot/system/manager/process_config.py

@@ -170,6 +170,7 @@ procs += [
   # Models
   PythonProcess("models_manager", "openpilot.sunnypilot.models.manager", only_offroad),
   NativeProcess("modeld_tinygrad", "openpilot/sunnypilot/modeld_v2", ["./modeld"], and_(only_onroad, is_tinygrad_model)),
+  PythonProcess("modeld_warmup", "openpilot.sunnypilot.modeld_v2.warmup", only_offroad),
 
   # Backup
   PythonProcess("backup_manager", "openpilot.sunnypilot.sunnylink.backups.manager", and_(only_offroad, sunnylink_ready_shim)),

release/ci/model_generator.py

@@ -54,18 +54,10 @@ def create_pkl_name(full_name: str) -> str:
 
 
 def _hash_pkl(pkl_path: Path) -> str:
-  manifest = Path(f"{pkl_path}.chunkmanifest")
-  if manifest.exists():
-    num_chunks = int(manifest.read_text().strip())
-    paths = [Path(f"{pkl_path}.chunk{i + 1:02d}of{num_chunks:02d}") for i in range(num_chunks)]
-  else:
-    paths = [pkl_path]
-
   digest = hashlib.sha256()
-  for path in paths:
-    with path.open('rb') as f:
-      while block := f.read(1024 * 1024):
-        digest.update(block)
+  with pkl_path.open('rb') as f:
+    while block := f.read(1024 * 1024):
+      digest.update(block)
   return digest.hexdigest()
 
 
@@ -74,22 +66,9 @@ def _find_driving_pkl(output_path: Path) -> Path | None:
     matches = sorted(output_path.glob(pattern))
     if matches:
       return matches[0]
-  for pattern in ('*driving_tinygrad.pkl.chunkmanifest', '*driving_*_tinygrad.pkl.chunkmanifest'):
-    matches = sorted(output_path.glob(pattern))
-    if matches:
-      return Path(str(matches[0]).removesuffix('.chunkmanifest'))
   return None
 
 
-def _rename_pkl_with_chunks(old_pkl: Path, new_pkl: Path) -> Path:
-  manifest = Path(f"{old_pkl}.chunkmanifest")
-  if manifest.exists():
-    for f in sorted(old_pkl.parent.glob(f"{old_pkl.name}.chunk*")):
-      f.rename(old_pkl.parent / f.name.replace(old_pkl.name, new_pkl.name, 1))
-    return new_pkl
-  return old_pkl.rename(new_pkl)
-
-
 def _hash_onnx_files(model_dir: Path) -> str | None:
   onnx_files = [f for f in sorted(model_dir.glob("*.onnx")) if f.stat().st_size > 1024]
   if not onnx_files:
@@ -102,22 +81,9 @@ def _hash_onnx_files(model_dir: Path) -> str | None:
   return digest.hexdigest()
 
 
-def generate_chunked_model(driving_pkl: Path) -> dict:
+def generate_model(driving_pkl: Path) -> dict:
   tinygrad_hash = _hash_pkl(driving_pkl)
 
-  chunks_config = []
-  manifest_file = Path(f"{driving_pkl}.chunkmanifest")
-  if manifest_file.exists():
-    num_chunks = int(manifest_file.read_text().strip())
-    for i in range(num_chunks):
-      chunk_path = Path(f"{driving_pkl}.chunk{i + 1:02d}of{num_chunks:02d}")
-      if chunk_path.exists():
-        chunk_hash = hashlib.sha256(chunk_path.read_bytes()).hexdigest()
-        chunks_config.append({
-          "file_name": chunk_path.name,
-          "sha256": chunk_hash
-        })
-
   artifact_data = {
     "file_name": driving_pkl.name,
     "download_uri": {
@@ -126,9 +92,6 @@ def generate_chunked_model(driving_pkl: Path) -> dict:
     }
   }
 
-  if chunks_config:
-    artifact_data["chunks"] = chunks_config
-
   return {
     "type": "chunked",
     "artifact": artifact_data,
@@ -193,11 +156,11 @@ if __name__ == "__main__":
   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)
+      _driving_pkl = _driving_pkl.rename(new_pkl)
     else:
       _driving_pkl = new_pkl
 
-  _model_metadata = generate_chunked_model(_driving_pkl)
+  _model_metadata = generate_model(_driving_pkl)
   _onnx_sha256 = args.onnx_sha256 or _hash_onnx_files(Path(args.model_dir))
   create_metadata_json([_model_metadata], _output_dir, args.custom_name, _short_name, args.is_20hz, args.upstream_branch,
                        onnx_sha256=_onnx_sha256, is_big=is_big)

release/ci/upload_default_model.py

@@ -79,8 +79,6 @@ def main():
   artifact = bundle['models'][0]['artifact']
   hf_base = f"https://huggingface.co/datasets/{args.hf_repo}/resolve/main/{args.hf_defaults_path}/{folder_name}"
   artifact['download_uri']['url'] = f"{hf_base}/{artifact['file_name']}"
-  for chunk in artifact.get('chunks', []):
-    chunk['url'] = f"{hf_base}/{chunk['file_name']}"
 
   print(f"Uploading model to {args.hf_defaults_path}/{folder_name}/")
   api.upload_folder(