What are the VRAM requirements and recommended GPU setup for running FantasyWorld inference?

[JianHaoYin]

Thanks for releasing FantasyWorld — this unified video + 3D world modeling framework looks very promising! I am planning to run inference with your provided code and have a few questions about hardware requirements.

1. VRAM requirement

• What is the approximate GPU memory (VRAM) needed to run a single inference pass with the default model (e.g., frozen Wan2.1-based backbones + geometry branch), especially for a typical case like ~80 frames at 480p?
• Do you have any measurements of peak GPU memory usage during inference?

2. Recommended inference GPU configuration

• What would you consider the best recommended GPU setup for stable inference with FantasyWorld?
• For example, is a single high-memory card (e.g., 48 GB+) sufficient, or do you recommend multi-GPU (e.g., 4×/8× 24 GB GPUs)?
• Are there preferred strategies such as model sharding / tensor parallelism / DeepSpeed inference to distribute across GPUs?

Thank you very much for your time and help!

[pyramidpoint]

Thanks for releasing FantasyWorld — this unified video + 3D world modeling framework looks very promising! I am planning to run inference with your provided code and have a few questions about hardware requirements.

1. VRAM requirement

• What is the approximate GPU memory (VRAM) needed to run a single inference pass with the default model (e.g., frozen Wan2.1-based backbones + geometry branch), especially for a typical case like ~80 frames at 480p?
• Do you have any measurements of peak GPU memory usage during inference?

2. Recommended inference GPU configuration

• What would you consider the best recommended GPU setup for stable inference with FantasyWorld?
• For example, is a single high-memory card (e.g., 48 GB+) sufficient, or do you recommend multi-GPU (e.g., 4×/8× 24 GB GPUs)?
• Are there preferred strategies such as model sharding / tensor parallelism / DeepSpeed inference to distribute across GPUs?

Thank you very much for your time and help!

Thank you for your interest in FantasyWorld!

Here are the details regarding your questions about hardware requirements and inference configurations:

We currently have two versions of the model available, each with its own GPU memory footprint:

FantasyWorld-Wan2.1-I2V-14B-480P: A single inference pass requires approximately 58.24 GB of VRAM. We have successfully tested and run this version on both NVIDIA H20 and A100 GPUs.

FantasyWorld-Wan2.2-Fun-A14B-Control-Camera: This version is much larger and has a peak GPU memory usage of 135.87 GB during inference.

Regarding distribution across multiple GPUs: We currently do not have any inference parallelism strategies implemented (such as model sharding, tensor parallelism, or DeepSpeed). The models need to be loaded and run without built-in distribution methods.

I hope this helps you plan your setup! Please feel free to reach out if you have any further questions.

[LeeKeyu]

Thanks for releasing FantasyWorld — this unified video + 3D world modeling framework looks very promising! I am planning to run inference with your provided code and have a few questions about hardware requirements.
1. VRAM requirement

• What is the approximate GPU memory (VRAM) needed to run a single inference pass with the default model (e.g., frozen Wan2.1-based backbones + geometry branch), especially for a typical case like ~80 frames at 480p?
• Do you have any measurements of peak GPU memory usage during inference?

2. Recommended inference GPU configuration

• What would you consider the best recommended GPU setup for stable inference with FantasyWorld?
• For example, is a single high-memory card (e.g., 48 GB+) sufficient, or do you recommend multi-GPU (e.g., 4×/8× 24 GB GPUs)?
• Are there preferred strategies such as model sharding / tensor parallelism / DeepSpeed inference to distribute across GPUs?

Thank you very much for your time and help!

Thank you for your interest in FantasyWorld!

Here are the details regarding your questions about hardware requirements and inference configurations:

We currently have two versions of the model available, each with its own GPU memory footprint:

FantasyWorld-Wan2.1-I2V-14B-480P: A single inference pass requires approximately 58.24 GB of VRAM. We have successfully tested and run this version on both NVIDIA H20 and A100 GPUs.

FantasyWorld-Wan2.2-Fun-A14B-Control-Camera: This version is much larger and has a peak GPU memory usage of 135.87 GB during inference.

Regarding distribution across multiple GPUs: We currently do not have any inference parallelism strategies implemented (such as model sharding, tensor parallelism, or DeepSpeed). The models need to be loaded and run without built-in distribution methods.

I hope this helps you plan your setup! Please feel free to reach out if you have any further questions.

@pyramidpoint Hi, may I ask about the minimum GPU requirements to train FantasyWorld?