Request for timeline on 4-GPU, 4-step inference version release

[gamechanger-creator]

Hello LiveAvatar team,

First of all, thank you for open-sourcing this impressive work. The real-time streaming performance and engineering optimizations in LiveAvatar are very exciting and highly valuable for real-time interactive avatar applications.

I noticed in the README that you mentioned plans to support a 4-GPU setup with 4-step inference and to remove the dependency on additional single-GPU VAE parallelism through integration of the LightX2V VAE component. This would make deployment significantly more accessible for many users who do not have access to 5×H800 GPUs.

I wanted to ask if you could share any estimated timeline or progress update on:

The release of the 4-GPU inference pipeline

Support for the 4-step inference version on 4 GPUs

Integration of the LightX2V VAE component

Any planned optimizations to improve accessibility on smaller multi-GPU setups

This feature would greatly help teams like ours deploy LiveAvatar in production environments with more practical hardware configurations.

Thank you again for your excellent work and for contributing this to the community. Looking forward to your update.

[Yubo-Shankui]

:) We might not release it very soon. However, adjusting timestep_shift with the current version's weights for a three-step process may help.

[gamechanger-creator]

Hi @Yubo-Shankui,

Thank you for the suggestion regarding adjusting timestep_shift for a three-step process.

Just to clarify — if we adjust timestep_shift using the current released weights, would that allow us to:

Run a 3-step inference setup on 3 GPUs reliably?

And similarly, apply the same approach to achieve a 4-step inference configuration on 4 GPUs?

Or are there additional architectural changes required (e.g., pipeline parallelism modifications, VAE handling, scheduler constraints) beyond just tuning timestep_shift?

Thanks again for your guidance — this would really help us evaluate deployment feasibility on smaller multi-GPU setups.

[Yubo-Shankui]

Please also change --num_gpus_dit and --sample_steps to 3, and set --nproc_per_node to 4.