qwen3_vl

Model type

qwen3_vl

Class

Qwen3VL3ModelCausalLMModel

Task

qwen-vl

Source

models/qwen_vl.py

Description

Qwen3-VL vision-language model (3-model split).

Builds three separate ONNX models for onnxruntime-genai:

  • decoder: text decoder taking inputs_embeds (interleaved MRoPE)

  • vision_encoder: packed-attention ViT outputting merged features

  • embedding: token embedding + image feature fusion

.. note:: When deepstack_visual_indexes is set, DeepStack intermediate features are packed into image_features by the vision encoder. The embedding model scatters and flattens them into the generic rank-3 per_layer_inputs contract consumed by ORT GenAI, and the decoder restores and injects one map into each of its first D layers. Models without DeepStack indices are unaffected.

Usage

mobius build --model <MODEL_ID> output_dir/

from mobius import build

model = build("<MODEL_ID>")