[Ray] Add Support for Disaggregating VAE and DiT #422
Merged
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR introduces support for disaggregating the VAE and DiT using Ray.
We have introduced a new argument --vae_parallel_size and --dit_parallel_size. When vae_parallel_size is set to 0, the VAE and DiT remain combined and are not separated. If vae_parallel_size is set to 1 or greater, the VAE is placed in a separate parallel Ray worker, enabling independent parallelism.
ray_world_size = vae_parallel_size + dit_parallel_size
dit_parallel_size = (data_parallel_degree *
classifier_free_guidance_degree *
sequence_parallel_degree *
pipeline_parallel_degree *
tensor_parallel_degree)
Currently, only FluxPipeline is supported for testing. Support for additional pipelines is in progress.
The separation of loading VAE and DiT is underway. For now, both modules are fully loaded within each Ray worker.
It is necessary to use the latest DistVAE.
Each configuration was tested three times.