OOM issues with Wan2.1-T2V-14B on A100 80GB and 120GB CPU RAM

[wxh6688]

Thank you very much for your excellent work and for making the code publicly available.

I have been trying to reproduce your results on an NVIDIA A100 machine and apply the framework to the Wan2.1-T2V-14B model. However, I encountered some memory-related issues and would appreciate any guidance you could provide.

My hardware configuration is:

• GPU: NVIDIA A100 80GB
• CPU RAM: 120GB

Using the following settings:

size: 832x480
frame: 81
offload_model: true

offloading:
global_disable_offloading: false
attn.out_cache: true
attn.indices: true
attn.counts: true
text_encoders: true

I found that the CPU memory usage exceeded the available 120GB RAM.

When I changed:

attn.out_cache: false

the CPU memory usage became manageable, but then the GPU ran out of memory.

Eventually, I was able to run the model successfully with the following configuration:

offload_model: false

offloading:
global_disable_offloading: false
attn.out_cache: true
attn.indices: false
attn.counts: false
text_encoders: false

and by modifying LARGE_BUF_SIZE in offloaded_tensor.py to: 1 * 40 * 33000 * 128 * torch.finfo(torch.bfloat16).bits // 8

I noticed that your experiments were conducted at a much larger setting (720×1280 resolution and 121 frames). Therefore, I am wondering whether you would be willing to share the hardware specifications used in your experiments, particularly:

• GPU model and VRAM capacity
• CPU RAM size

This information would be very helpful for understanding the memory requirements and determining whether my issues are caused by hardware limitations or configuration differences.

Thank you again for your time and for the great work.