How qwenlm/qwen2-vl works
Qwen3-VL is a vision-language model release repository with supporting inference utilities, a chat demo, fine-tuning code, task cookbooks, and benchmark evaluation programs. The user-facing workflow is to submit text and image/video content to the demo, which prepares a model request and presents generated output. Fine-tuning and evaluation are separate workflows: training prepares multimodal examples and runs a trainer; benchmark programs prepare datasets, run inference, and score results. Connections reflect the README and indexed imports; unsampled implementation details are not inferred.
Main entry points: dataset_utils.py MMMU dataset · eval_utils.py MMMU scoring · common_utils.py MMMU helpers · train_qwen.py Training entry
Components 19
Inference demo
- Chat demo
web_demo_mm.py
Vision utilities
- Vision processing
qwen-vl-utils/src/qwen_vl_utils/vision_process.py
Fine-tuning
- Training entry
qwen-vl-finetune/qwenvl/train/train_qwen.py - Data processing
qwen-vl-finetune/qwenvl/data/data_processor.py - 2D position encoding
qwen-vl-finetune/qwenvl/data/rope2d.py - Training arguments
qwen-vl-finetune/qwenvl/train/argument.py - Trainer
qwen-vl-finetune/qwenvl/train/trainer.py
Benchmark evaluation
- MMMU inference
evaluation/mmmu/run_mmmu.py - MMMU dataset
evaluation/mmmu/dataset_utils.py - MMMU helpers
evaluation/mmmu/common_utils.py - MMMU scoring
evaluation/mmmu/eval_utils.py - MathVision inference
evaluation/MathVision/run_mathv.py - MathVision scoring
evaluation/MathVision/eval_utils.py - ODinW evaluation
evaluation/ODinW-13/run_odinw.py - RealWorldQA inference
evaluation/RealWorldQA/run_realworldqa.py - RealWorldQA scoring
evaluation/RealWorldQA/eval_utils.py - VideoMME inference
evaluation/VideoMME/run_videomme.py
Other components
- Demo user
- Task cookbooks
cookbooks
Connections 21
- Demo user → Chat demo: submits content
README.md - Chat demo → Vision processing: uses vision utilities
web_demo_mm.py - Training entry → Data processing: uses
qwen-vl-finetune/qwenvl/train/train_qwen.py - Training entry → Trainer: uses
qwen-vl-finetune/qwenvl/train/train_qwen.py - Training entry → Training arguments: uses
qwen-vl-finetune/qwenvl/train/train_qwen.py - Data processing → 2D position encoding: uses
qwen-vl-finetune/qwenvl/data/data_processor.py - MMMU inference → MMMU dataset: loads dataset
evaluation/mmmu/run_mmmu.py - MMMU inference → MMMU scoring: uses scoring
evaluation/mmmu/run_mmmu.py - MMMU dataset → MMMU helpers: uses helpers
evaluation/mmmu/dataset_utils.py - MMMU scoring → MMMU helpers: uses helpers
evaluation/mmmu/eval_utils.py - MathVision inference → MMMU dataset: uses dataset helpers
evaluation/MathVision/run_mathv.py - MathVision inference → MMMU scoring: uses scoring helpers
evaluation/MathVision/run_mathv.py - MathVision scoring → MMMU helpers: uses helpers
evaluation/MathVision/eval_utils.py - ODinW evaluation → MMMU dataset: uses dataset helpers
evaluation/ODinW-13/run_odinw.py - ODinW evaluation → MMMU scoring: uses scoring helpers
evaluation/ODinW-13/run_odinw.py - RealWorldQA inference → MMMU dataset: uses dataset helpers
evaluation/RealWorldQA/run_realworldqa.py - RealWorldQA inference → MMMU scoring: uses scoring helpers
evaluation/RealWorldQA/run_realworldqa.py - RealWorldQA scoring → MMMU helpers: uses helpers
evaluation/RealWorldQA/eval_utils.py - VideoMME inference → MMMU dataset: uses dataset helpers
evaluation/VideoMME/run_videomme.py - VideoMME inference → MMMU scoring: uses scoring helpers
evaluation/VideoMME/run_videomme.py - Task cookbooks → Demo user: demonstrates tasks
README.md