Instructions to use gatilin/sensenova-loop-9b-vision-encoder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use gatilin/sensenova-loop-9b-vision-encoder with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("feature-extraction", model="gatilin/sensenova-loop-9b-vision-encoder", trust_remote_code=True)# pip install -U transformers accelerate # Load model directly from transformers import AutoModelForCausalLM model = AutoModelForCausalLM.from_pretrained("gatilin/sensenova-loop-9b-vision-encoder", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
SenseNova-Loop-9B Vision Encoder
This repository contains the visual tower and visual merger extracted from SenseNova-Loop-9B. It does not contain the language model or a complete multimodal assistant. All original model.visual.* tensor names are preserved.
- Architecture: Qwen3.5 vision encoder, 27 transformer blocks, 16 attention heads, hidden dimension 1,152, spatial patch size 16 and temporal patch size 2.
- The included merger combines spatial 2 ร 2 patches and projects to dimension 4,096.
- 333 tensors, 456,010,480 parameters, BF16, about 912 MB of tensor data.
- Weights SHA-256:
a657c406083ff2cb08b562e4eb819206498b51401e2554d01146405848a309a7.
Example
Install the packages in requirements.txt and download this repository. component.py loads the upstream Qwen3.5 visual implementation, strips the checkpoint prefix, and strictly checks every weight. The example uses the original image processor settings. The output includes per-patch hidden states and merged features for the language model.
from pathlib import Path
import torch
from PIL import Image
from transformers import Qwen2VLImageProcessor
from component import load_vision_model
root = Path('.') # local copy of this repository
model = load_vision_model(root, dtype=torch.float32, device='cpu')
processor = Qwen2VLImageProcessor.from_pretrained(root)
inputs = processor(images=Image.open('image.jpg').convert('RGB'), return_tensors='pt')
with torch.inference_mode():
outputs = model(inputs['pixel_values'], grid_thw=inputs['image_grid_thw'])
print(outputs.last_hidden_state.shape, outputs.pooler_output.shape)
The standalone bundle is intended for visual features and attention research. It does not implement text generation or reproduce the original model's end-to-end multimodal predictions.
Provenance and verification
The source Hugging Face revision is 3e1ba6e8ac98090f9f7c03a55f26269eb3850fb6. Its gated weight files were inaccessible to the account at extraction time, so the checkpoint ranges were downloaded from the official SenseNova ModelScope mirror, using the file revisions recorded in extraction.json. Source config, index and processor/code files match the Hugging Face Git blob identifiers; the two providers' declared full-shard sizes and SHA-256 values match. The extracted tensor payload was compared byte for byte with the downloaded mirror ranges. The complete original shards were not downloaded and hashed locally, and Hugging Face shard headers were not independently compared. Details are in source_verification.json, tensor_mapping.json and extraction.json.
validation.json records strict loading and CPU float32 forwards on two COCO128 images. For each image, explicit attention was checked against native SDPA at layers 0, 13 and 26. This smoke test uses 224 ร 224 inputs and is not an end-to-end equivalence test of the full original model. See validate_local_reference.py to reproduce it.
License
Apache-2.0, as indicated by the upstream model card. Review the upstream model card for attribution and any applicable terms.
- Downloads last month
- 9
Model tree for gatilin/sensenova-loop-9b-vision-encoder
Base model
sensenova/SenseNova-Loop-9B