This delivery adapts the official philschmid/layoutlm-funsd checkpoint for a real Ascend NPU compatibility smoke. The official model page and upstream LayoutLM source are:
The verified Hugging Face revision is 5dc97b05b27e5efed5f331a2216726dd7b338e43. The model card API does not declare a license, so this delivery records Unknown rather than inferring one. The official LayoutLMForTokenClassification model was strictly loaded on npu:0. Synthetic token and bounding-box inputs with shapes [1, 16] and [1, 16, 4] produced finite bfloat16 logits with shape [1, 16, 7]; the real NPU forward completed in about 0.103 seconds.
This is an execution compatibility gate only. It does not claim FUNSD token-classification quality or full document-task precision. The official weight and tokenizer files are excluded from Git and must be downloaded into the model directory before running:
python inference.py --model-dir /tmp/model-agent/layoutlm --output validation/npu-runtime.jsonThe verified weight size and SHA-256 are recorded in model-lock.json.


