凤
weixin_44093752/chandra-ocr-2-ascend-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

Chandra OCR 2 Ascend NPU smoke adaptation

This delivery tracks the live datalab-to/chandra-ocr-2 checkpoint and the official Datalab implementation:

  • model: https://huggingface.co/datalab-to/chandra-ocr-2
  • revision: af93b47dba1b47b6640c86ccf487ed2260ab9a09
  • source: https://github.com/datalab-to/chandra
  • source revision: d4f7467435aa4137d9539f000ddf0b7ced3eb43f
  • weight: model.safetensors, 10,591,220,088 bytes
  • SHA-256: 0804568be9f099d6479fad9ed77a4da4611f3c1e7bc6e009af7dce45e8aa3847
  • model class: Qwen3_5ForConditionalGeneration

inference.py loads the official Transformers model locally, creates a deterministic white document image, and generates four OCR tokens on npu:0. The verified Ascend910 run loaded 4,539,265,536 parameters and produced finite token IDs. The acceptance gate is real NPU execution plus finite output, as requested for this competition pass; it does not claim a complete OCR quality or CPU/NPU precision gate.

The model card declares openrail; the weights use the model card's modified OpenRAIL-M terms. Review the upstream license before production use.

Run after downloading the checkpoint to a local directory:

python inference.py --model-path /tmp/model-agent/chandra-ocr-2 --device npu --output-dir validation

Model Agent CPU/NPU workflow

Ascend NPU device call

NPU smoke result