This repository documents a verified full forward pass of the official
Addaoud/DeepPlant-GEP checkpoint on Ascend NPU (npu:0). It is an inference
adaptation only; no weights are modified, replaced, trained, or fine-tuned.
Addaoud/DeepPlant-GEP, Phase 4 row 86, record recvsfrbvP5eab57d02f6e9bfee46c98c8a0cbc1b835095898a3b675456915b0b6d09b444b43ab1ef803a0b129a9a1938d9d92ccaae4d7592def90bbdd95dec527f1f54b101d940a758a3da3aaf0c8The model Hub configuration has 2,935 GEP targets. The source dataclass uses a one-item list for that value, so the runner normalizes only this serialization difference. Nine checkpoint keys are mapped one-to-one to equivalent nested module names; every mapped tensor shape is checked and loading must end with zero missing and zero unexpected keys.
Run inference.py with the official model directory, official source directory,
and the pinned Git repository directory. The runner validates the checkpoint
hash, compares all 20 executable src/ Git blobs to the fixed upstream commit,
places model, input and output on npu:0, and checks finite output of shape
[2935]. PyTorch Transformer fast-path is explicitly disabled because this
installed torch_npu stack routes that fused fast-path through CPU; constituent
operations remain on NPU and any fallback invalidates the evidence.
The exact evidence is in evidence/result.json; screenshots record the active
Model Agent workflow, NPU device path, and final result without altering the
official model weights.