# Eagle | NVIDIA Research Projects | Knowledge Base | Instagit

Eagle: Frontier Vision-Language Models with Data-Centric Strategies

GitHub Stars: 2.8k

Repository: https://github.com/NVlabs/Eagle

---

## Articles

### [How to Configure FlashAttention for Eagle Inference on Different GPUs](/NVlabs/Eagle/how-to-configure-flashattention-for-eagle-inference-on-different-gpus)

Learn how to configure FlashAttention for Eagle inference across diverse GPUs. Install the right FlashAttention 2 wheel and enable optimized kernels for faster performance.

- Tags: how-to-guide
- Published: 2026-06-28

### [Eagle Tokenizer Configuration for Multimodal Inputs: A Complete Guide](/NVlabs/Eagle/what-tokenizer-configuration-does-eagle-use-for-multimodal-inputs)

Discover the Eagle tokenizer configuration for multimodal inputs. Learn how Eagle uses Hugging Face AutoTokenizer with special image tokens and padding for vision-language encoding.

- Tags: how-to-guide
- Published: 2026-06-28

### [How LocateAnything Performs Zero-Shot Object Detection in the Wild: A Technical Deep Dive](/NVlabs/Eagle/how-does-locateanything-perform-zero-shot-object-detection-in-the-wild)

Unlock zero-shot object detection with LocateAnything. Discover how this unified vision-language model achieves precise detection in the wild without fine-tuning. Learn the technical details.

- Tags: deep-dive
- Published: 2026-06-28

### [How to Add a Custom Evaluation Task to Eagle's LMMS-Eval Framework](/NVlabs/Eagle/how-to-add-a-custom-evaluation-task-to-eagles-lmms_eval-framework)

Easily add custom evaluation tasks to Eagle's LMMS-Eval framework. Create YAML and Python files in a new folder for automatic discovery without core code changes.

- Tags: how-to-guide
- Published: 2026-06-28

### [Memory Optimization Techniques in Eagle for Training Large Multimodal Models](/NVlabs/Eagle/what-memory-optimization-techniques-does-eagle-use-for-training-large-models)

Discover Eagle's memory optimization techniques like FlashAttention and gradient checkpointing for efficient large model training. Train 8B models on a single GPU.

- Tags: deep-dive
- Published: 2026-06-28

### [How Eagle's Vision-Language Projector Handles Feature Fusion from Multiple Encoders](/NVlabs/Eagle/how-does-the-vision-language-projector-handle-feature-fusion-from-multiple-encoders)

Discover how Eagle fuses features from multiple vision encoders. Learn about its configurable linear layer or MLP projection into the LLM's hidden space.

- Tags: internals
- Published: 2026-06-28

### [How to Debug Vision Tower Loading Issues in Eagle Training: A Complete Guide](/NVlabs/Eagle/how-to-debug-vision-tower-loading-issues-in-eagle-training)

Debug Eagle training vision tower loading issues. Verify paths inspect keys check config and run a dummy pass. Master NVlabs/Eagle troubleshooting for efficient AI development.

- Tags: how-to-guide
- Published: 2026-06-28

### [Understanding Spatial and Flat Patch Merge Types in Eagle](/NVlabs/Eagle/what-is-the-difference-between-spatial-and-flat-patch-merge-types-in-eagle)

Discover the difference between spatial and flat patch merge types in Eagle models. Learn how mm_patch_merge_type configures 1D vs 2D patch processing for enhanced visual understanding.

- Tags: deep-dive
- Published: 2026-06-28

### [How Eagle Handles Multi-Image Input in a Single Forward Pass: Architecture and Implementation](/NVlabs/Eagle/how-does-eagle-handle-multi-image-input-in-a-single-forward-pass)

Discover how Eagle handles multi-image input in a single forward pass. Learn its architecture and implementation for efficient image processing and feature integration.

- Tags: architecture
- Published: 2026-06-28

### [How to Integrate Eagle with TensorRT for Optimized Inference](/NVlabs/Eagle/how-to-integrate-eagle-with-tensorrt-for-optimized-inference)

Integrate Eagle with TensorRT for optimized inference by exporting to ONNX, building a TensorRT engine, and replacing PyTorch execution.

- Tags: how-to-guide
- Published: 2026-06-28

### [Data Post-Training Strategies in Eagle 2 for Frontier Vision-Language Models](/NVlabs/Eagle/what-data-post-training-strategies-are-used-in-eagle-2-for-frontier-vlms)

Discover Eagle 2.5's data post-training strategies, including Progressive Mixed Post-Training and Information-First Sampling, that expand context windows to 128K tokens while maintaining visual fidelity.

- Tags: data-post-training-strategies
- Published: 2026-06-28

### [Eagle 2.5 Long-Context Training Strategy: Processing 28K Tokens Efficiently](/NVlabs/Eagle/what-is-eagle-2.5s-long-context-training-strategy-for-128k-tokens)

Discover Eagle 2.5's efficient long-context training strategy. Learn how its custom attention kernel and token packing process 28K tokens for stable, high-performance results.

- Tags: deep-dive
- Published: 2026-06-28

### [How to Convert and Consolidate Eagle Model Checkpoints for Deployment](/NVlabs/Eagle/how-to-convert-and-consolidate-eagle-model-checkpoints-for-deployment)

Easily convert and consolidate Eagle model checkpoints for deployment. Load sharded checkpoints, merge them automatically, and export a single file for production inference.

- Tags: how-to-guide
- Published: 2026-06-28

### [How to Evaluate Eagle on VLM Benchmarks Using lmms-eval: A Complete Guide](/NVlabs/Eagle/how-to-evaluate-eagle-on-vlm-benchmarks-using-lmms_eval)

Learn how to evaluate Eagle on VLM benchmarks with lmms-eval. Our guide covers loading checkpoints, preparing vision towers, and running batched inference for comprehensive vision-language task analysis.

- Tags: how-to-guide
- Published: 2026-06-28

### [Parallel Box Decoding in LocateAnything: From Sequential Tokens to Atomic Box Prediction](/NVlabs/Eagle/what-is-parallel-box-decoding-and-how-does-locateanything-use-it)

Discover Parallel Box Decoding in LocateAnything. This method predicts entire bounding boxes at once for 2x-6x faster decoding than token-by-token approaches. Learn more!

- Tags: deep-dive
- Published: 2026-06-28

### [How Eagle Handles Variable-Sized Images in Its Encoding Pipeline](/NVlabs/Eagle/how-does-eagles-image-encoding-pipeline-handle-variable-sized-images)

Discover how Eagle's encoding pipeline processes variable-sized images, using dynamic routines for padding or optimal grid resolutions based on image aspect ratio.

- Tags: internals
- Published: 2026-06-28

### [Supported LLM Backbones in Eagle and How to Switch Between Them](/NVlabs/Eagle/what-are-supported-llm-backbones-in-eagle-and-how-to-switch)

Discover supported LLM backbones like LLaMA OPT Mistral and Falcon in Eagle. Easily switch between models using command-line arguments without code changes.

- Tags: api-reference
- Published: 2026-06-28

### [How to Run Inference with Eagle Using the Gradio Demo Interface](/NVlabs/Eagle/how-to-run-inference-with-eagle-using-the-gradio-demo-interface)

Easily run inference with the Eagle multimodal model using its Gradio demo interface. Interact with a pretrained model for chat-style conversations. Get started now.

- Tags: how-to-guide
- Published: 2026-06-28

### [How to Fine-Tune Eagle Models with LoRA for Efficient Training](/NVlabs/Eagle/how-to-fine-tune-eagle-models-with-lora-for-efficient-training)

Fine-tune Eagle models efficiently with LoRA. This guide shows how to enable LoRA adapters and configure ranks to train faster and save GPU memory without altering original weights.

- Tags: tutorial
- Published: 2026-06-28

### [How to Add a New Custom Vision Encoder to Eagle's Architecture](/NVlabs/Eagle/how-to-add-a-new-custom-vision-encoder-to-eagles-architecture)

Learn to add a custom vision encoder to Eagle's architecture. Implement a tower wrapper, register it, and update CLI configuration for seamless integration. Explore NVlabs/Eagle.

- Tags: how-to-guide
- Published: 2026-06-28

### [What Is AnyRes and How It Handles Multiple Image Resolutions in Eagle VLM](/NVlabs/Eagle/what-is-anyres-and-how-does-it-handle-multiple-image-resolutions-in-eagle-vlm)

Discover AnyRes, a PyTorch dataset wrapper used by Eagle VLM to train on diverse image resolutions. Learn how it manages variable-length lists and token-level flags for efficient multi-resolution handling.

- Tags: tutorial
- Published: 2026-06-28

### [How to Configure the Multimodal Projector for Different Vision Encoders in Eagle](/NVlabs/Eagle/how-to-configure-multimodal-projector-for-different-vision-encoders-in-eagle)

Configure Eagle's multimodal projector for various vision encoders. Learn to set mm_projector_type and mm_hidden_size to optimize your model's performance.

- Tags: how-to-guide
- Published: 2026-06-28

### [How Eagle's Mixture-of-Encoders Combines Multiple Vision Encoders](/NVlabs/Eagle/how-does-eagles-mixture-of-encoders-work-with-multiple-vision-encoders)

Learn how Eagle's mixture-of-encoders aggregates heterogeneous vision backbones. This strategy dynamically loads multiple encoders, aligns spatial resolutions, and concatenates features for a unified visual representation.

- Tags: internals
- Published: 2026-06-28