Bridging the Last Mile: Deploying Hummingbird-XT for Efficient Video Generation on AMD Consumer-Grade Platforms

🔍 Overview

This repository presents an efficient acceleration pipeline for Diffusion Transformer (DiT) based video generation models, optimized for AMD client-grade GPUs, including Navi48 dGPUs and Strix Halo iGPUs.

Built upon this pipeline, we introduce Hummingbird-XT, a new family of DiT-based text-to-video models derived from Wan2.2-5B, achieving high-quality video generation with significantly reduced inference cost.Additionally, to further extend the length of generated videos, we introduce Hummingbird-XTX, an efficient autoregressive model for long-video generation based on Wan-2.1-1.3B, which is capable of generating long videos.

Hummingbird-XT Text-to-Video Showcases

Hummingbird-XT Image-to-Video Showcases

Hummingbird-XTX 20s videos Showcases

📝 News

• [2026.07.13]: 🔥🔥We present a ROCm-optimized implementation of video sparse attention (VSA) that delivers 1.37x and 1.33x speedups in the forward and backward passes, respectively, over the original FastVideo implementation on AMD Instinct MI325X GPUs.
• [2026.02.10]: 🔥🔥Based on the Wan 2.2 5B model, we trained a 5B version of the Hummingbird-XTX long-video generation model, and have now open-sourced all the code and weights.
• [2026.01.09]: 🔥🔥Release the full code and pre-trained weight of HummingbirdXT!
• [2026.01.08]: 🔥🔥Release our Blog: Bridging the Last Mile: Deploying Hummingbird-XT for Efficient Video Generation on AMD Consumer-Grade Platforms !

🧬 Models

Hummingbird-XT

• DiT-based text-to-video model built upon Wan2.2-5B
• Optimized for few-step inference
• Designed for efficient deployment on AMD GPUs
• Maintains competitive visual quality compared to full-step baselines

Hummingbird-XTX (Long Video Extension)

• Extends Hummingbird-XT to support efficient long video generation
• Improves temporal consistency across extended sequences
• Suitable for long-form generation scenarios

Lightweight VAE Decoder

• Lightweight and efficient VAE decoder

• Introduces 3D DW Conv to reduce redundancy in original 3D Conv

• Achieves 14x speedup for decoding and reduce memory usage by 4x

• Preserves competitive reconstruction and generation quality

⚙️ Installation

Clone this Repo:

git clone https://github.com/AMD-AGI/HummingbirdXT.git
cd HummingbirdXT

Option 1: Conda Environment

conda create -n hummingbirdxt python=3.10
conda activate hummingbirdxt
pip install -r requirements.txt

For rocm flash-attn, you can install it by this link.

git clone https://github.com/ROCm/flash-attention.git
cd flash-attention
python setup.py install

Option 2: Docker

You can download our pre-built Docker image for better reproducibility:

docker pull panisobe/dmd_flash_image_2:latest

You can use docker run to run the image. For example:

docker run -it \
  --shm-size=900g \
  --name hm \
  --network host \
  --ipc=host \
  --cap-add=SYS_PTRACE \
  --security-opt seccomp=unconfined \
  --device=/dev/kfd \
  --device=/dev/dri \
  --group-add video \
  -e HSA_OVERRIDE_GFX_VERSION=11.0.0 \
  -v /home:/home \
  panisobe/dmd_flash_image_2_release:latest

🚀 Getting Started for Video Generation

You can download the weights for all our models from our models' huggingface: amd/HummingbirdXT.

Rocm-VSA

cd ./HummingbirdXT_VSA/train 
bash running_scripts/train/dmd.sh 

HummingbirdXT Video Generation

cd infer
bash run_t2v.sh # for text-to-video  task
bash run_i2v.sh # for image-to-video task

If you use docker, you should make sure diffusers==0.35.0. To use Lightweight VAE decoder in generation, add the below parameters after running command in run_t2v.sh and run_i2v.sh:

--vae_model ${VAE_PATH_ROOT}/wan22_v1_tiling_16_12 --t_block_size 16 --t_stride 12  

HummingbirdXTX Long Video Generation

For 5B model, you can run below commands:

cd long_video
bash infer.sh

For 1.3B model, you can run this commamd:

cd long_video
bash infer_1_3b.sh

🧪 Training & Implementation

HummingbirdXT Training

First you need to enter the train folder:

cd train

Step 1: Download the Teacher Model We use Wan2.2-TI2V-5B as the teacher model for step distillation.

pip install "huggingface_hub[hf_transfer]"
HF_HUB_ENABLE_HF_TRANSFER=1 --local-dir wan_models/Wan2.2-TI2V-5B

Step 2: Prepare Training Datasets We train our models on a mixture of large-scale video datasets, including: MagicData,OpenVid,HumanVid

Please update the dataset root paths in the corresponding CSV files to match your local storage layout. You can download the csv file from our models' huggingface: amd/HummingbirdXT.

Step 3: Launch Training Start the step distillation training using the provided script:

bash running_scripts/train/dmd.sh

The training pipeline demonstrates stable loss convergence across all models.

Reference Training Configuration: GPUs: 16 × AMD MI325, Iterations: 4000, Training time: ~48 hours.

HummingbirdXTX Training

Step 1: ODE Initialization(Optional)

cd long_video
bash train_ode.sh

Or you can directly download our trained ODE initialization weights from our models' huggingface: amd/HummingbirdXT, for the second stage of training.

Step 2: DMD Training

bash train_dmd.sh

📊 Experimental Results

Table 1. Quantitative results for the text-to-video task on VBench.

Table 2. Quantitative results for the image-to-video task on VBench.

Table 3. Runtime for generating a 121-frame video at 704×1280 resolution on server-grade (AMD Instinct™ MI300X and AMD Instinct™ MI325X GPU) and client-grade (Strix Halo and Navi48).

Table 4. Performance and efficiency comparison of different VAE decoders on AMD Instinct™ MI300X GPU.

Table 5. Quantitative results for long video generation on three benchmarks.

🤗Additional Resources

Huggingface model cards: AMD-HummingbirdXT

Full training code: AMD-AIG-AIMA/HummingbirdXT

Related work on diffusion models by the AMD team:

• AMD Hummingbird-0.9B: An Efficient Text-to-Video Diffusion Model with 4-Step Inferencing
• AMD Hummingbird Image to Video: A Lightweight Feedback-Driven Model for Efficient Image-to-Video Generation

Please refer to the following resources to get started with training on AMD ROCm™ software:

• Use the public PyTorch ROCm Docker images that enable optimized training performance out-of-the-box
• PyTorch Fully Sharded Data Parallel (FSDP) on AMD GPUs with ROCm — ROCm Blogs
• Accelerating Large Language Models with Flash Attention on AMD GPUs — ROCm Blogs

❤️ Acknowledgement

Our codebase builds on Wan 2.1, Wan 2.2, Self-Forcing, VideoX-Fun .Thanks the authors for sharing their awesome codebases!

📋 Citations

Feel free to cite our Hummingbird-XT models and give us a star⭐, if you find our work helpful. Thank you.