| |
| """CPU fp32 reference of BEVFormer-small (Autoware tensorrt_bevformer): the ground truth every PCC / output-agreement |
| gate compares against (PLAN.md 0.3 item 1; research/bevformer/SPEC.md section 6). Ported from the verified research |
| reference ``research/bevformer/scripts/bevformer_ref.py``. |
| |
| Layout (importable without ttnn; ``config`` / ``weights`` need numpy only, the rest imports torch): |
| |
| - ``config.py`` every constant of the deployed graph and the node, with its source line (cameras, classes, |
| geometry, image sizes, transformer sizes, post-processing defaults, SCA rebatch capacity). |
| - ``weights.py`` ``BEVFormerWeights``: the mirror's safetensors (or the official .pth) read as data with the |
| vendored ``ttaw.weights``; strict 1,001-tensor coverage; the backbone BatchNorms folded in float64 |
| exactly where the port folds them (``conv(module) -> ttaw.models.resnet.ConvParams``). |
| - ``ops.py`` the three TensorRT plugins (MSDA, DCNv2, RotateTRT) in torch fp32 + literal float64 CUDA-loop ports. |
| - ``model.py`` ``BEVFormerNet``: the network with per-module taps; ``forward`` (the deployed graph, geometry |
| inside) and ``forward_port`` (the TT port's decomposition with host geometry: rebatched SCA, |
| rotation by gather). |
| - ``pipeline.py`` ``BEVFormerReference``: frames from images + calibration + ego pose through ``host`` (the node's |
| pre-processing, temporal state, post-processing) and the network; stateless ``run_frame``. |
| - ``ort.py`` the ONNX Runtime oracle on the derived pure ONNX (research and tests only; onnxruntime is NOT a |
| runtime dependency of the image). |
| |
| The ``to_dict()`` of this reference's output on the shipped sample is stored as |
| ``samples/<sample stem>.reference.json``: ``server/smoke_test.py`` compares the served output with it. |
| """ |
|
|
