How nvlabs/eagle works
Eagle is a family of vision-language models spanning general image understanding (Eagle), long-context image/video understanding (Eagle 2.5), and grounding, detection, and pointing (LocateAnything). The user workflow is to submit an image or video with a prompt through a demo or inference entry point, have the model process visual and language inputs, and receive generated text or localized results; evaluation scripts assess model outputs on benchmark tasks. The graph includes principal model, training, and evaluation branches. Some subsystem paths and interactions are inferred only from the tree and README, so uncertain wiring is omitted.
Main entry points: modeling_eagle2_5_vl.py Eagle 2.5 VLM · modeling_locateanything.py LocateAnything Model · app.py Eagle 2.5 Demo · model_worker.py Model Worker
Components 20
User Interfaces
- Eagle 2.5 Demo
Eagle2_5/streamlit_demo/app.py - Demo Controller
Eagle2_5/streamlit_demo/controller.py - Model Worker
Eagle2_5/streamlit_demo/model_worker.py - LocateAnything Worker
Embodied/locateanything_worker.py
Model Runtime
- Eagle VLM
Eagle/eagle/model/eagle_arch.py - Eagle 2.5 VLM
Eagle2_5/eaglevl/model/eagle2_5/modeling_eagle2_5_vl.py - Vision Encoders
Eagle2_5/eaglevl/model/multimodal_encoder - Language Models
Eagle2_5/eaglevl/model/llama - LocateAnything Model
Embodied/eaglevl/model/locany/modeling_locateanything.py
Training
- Eagle 2.5 Training
Eagle2_5/eaglevl/train/eagle_2_5_vl_finetune.py - LocateAnything Training
Embodied/eaglevl/train/locany_finetune_magi_stream.py
Evaluation
- Multimodal Evaluator
Eagle/lmms_eval/evaluator.py - Benchmark Tasks
Eagle/lmms_eval/tasks - Grounding Evaluation
Embodied/evaluation/inference_grounding_ddp.py - Detection Evaluation
Embodied/evaluation/inference_detection_ddp.py - ScreenSpot Evaluation
Embodied/evaluation/inference_sspro_ddp.py - Evaluation Metrics
Embodied/evaluation/metrics
Other components
- User
- Image / Video Input
- Training Data
Connections 24
- User → Eagle 2.5 Demo: submits prompt
- Image / Video Input → Eagle 2.5 Demo: uploads media
- Eagle 2.5 Demo → Demo Controller: requests model
- Eagle 2.5 Demo → Model Worker: streams generation
- Demo Controller → Model Worker: dispatches request
- Model Worker → Eagle 2.5 VLM: generates response
- Eagle 2.5 VLM → Vision Encoders: encodes visuals
- Eagle 2.5 VLM → Language Models: processes language
- User → LocateAnything Worker: submits query
- Image / Video Input → LocateAnything Worker: provides image
- LocateAnything Worker → LocateAnything Model: generates grounding
- Training Data → Eagle 2.5 Training: provides examples
- Eagle 2.5 Training → Eagle 2.5 VLM: fine-tunes
- Training Data → LocateAnything Training: provides examples
- LocateAnything Training → LocateAnything Model: fine-tunes
- Eagle VLM → Vision Encoders: uses encoders
- Eagle 2.5 VLM → Multimodal Evaluator: supplies model
- Multimodal Evaluator → Benchmark Tasks: evaluates tasks
- LocateAnything Model → Grounding Evaluation: supplies predictions
- LocateAnything Model → Detection Evaluation: supplies predictions
- LocateAnything Model → ScreenSpot Evaluation: supplies predictions
- Grounding Evaluation → Evaluation Metrics: scores results
- Detection Evaluation → Evaluation Metrics: scores results
- ScreenSpot Evaluation → Evaluation Metrics: scores results