openEuler/siglip2_so400m_patch14_384

Model Card for SigLIP2 SO400M Patch14-384 (IB-Robot)

SigLIP2 (Sigmoid Loss for Language Image Pre-training 2) with SO400M backbone, patch size 14, 384x384 input. Produces 1152-dimensional L2-normalized image and text embeddings for vision-language matching. Packaged for the IB-Robot framework.

Repository Structure

• inference_manifest.json — deployment routing (schema v3)
• assets/model/ — HuggingFace model snapshot (safetensors, config, tokenizer)
• assets/adapter.json — deployment adapter config
• artifacts/ascend_310b/ — vision + text encoder OM
• artifacts/ascend_310p/ — vision + text encoder OM (aarch64)

Deployment Backends

Target	Backend	Runtime	Hardware
ascend_310b	ascend	ACL	Ascend 310B1
ascend_310p	ascend	ACL	Ascend 310P1
torch_cpu	torch	PyTorch	CPU
torch_cuda	torch	PyTorch	NVIDIA GPU

Each Ascend deployment runs a dual-encoder pipeline: vision + text.

Inputs: masked_images float32 [-1,3,384,384] NCHW, text_tokens int64 [-1,64], text_attention_mask int64 [-1,64] Outputs: image_embeddings float32 [-1,1152], text_embeddings float32 [-1,1152] Embedding space: L2-normalized, dimension 1152, google/siglip2-so400m-patch14-384@main

Source Model

This bundle's torch weights originate from the upstream SigLIP2 model:

• HuggingFace: google/siglip2-so400m-patch14-384

The Ascend OM artifacts were converted from these torch weights. Download script: scripts/download_perception_models.sh.

Citation

@inproceedings{siglip2,
  title     = {SigLIP 2: Multilingual Vision-Language Pre-Training with Improved Semantic Alignment},
  author    = {Zhai, Xiaohua and Mustafa, Basil and Kolesnikov, Alexander and Beyer, Lucas},
  booktitle = {arXiv preprint arXiv:2502.14795},
  year      = {2025}
}
@software{ib_robot,
  title   = {IB-Robot: Intelligence Boom Robot},
  url     = {https://gitcode.com/openeuler/IB_Robot},
  license = {Apache-2.0}
}