What's changed since 0.0.0beta5
-
Added distributed VAE encoding and decoding using spatial row sharding, initially for Wan and now generalized across supported image and video VAE families, by @nsakkine, @jjuvonen-amd, and @pds-amd in #10, #11, #12 and #14.
-
Added tile-parallel VAE execution with overlap-aware blending and distributed tile scheduling for high-resolution workloads by @pds-amd in #14.
-
Expanded parallel VAE support to Wan, Qwen-Image, HunyuanVideo, HunyuanVideo 1.5, LTX-2, LTX-2.3, Flux.2, Flux-Kontext, SD3, Z-Image, and compatible Diffusers AutoencoderKL models, by @nsakkine, @jjuvonen-amd, and @pds-amd in #10, #12 and #14.
-
Added public APIs for selecting row-sharded or tile-parallel VAE execution and integrating these strategies with xDiT by @pds-amd in #14.
-
Improved distributed correctness across uneven spatial splits, distributed normalization, causal state, padding, convolution halos, and rectangular inputs, by @nsakkine, @jjuvonen-amd, and @pds-amd in #10, #11, #12 and #14.
-
Added a reproducible benchmark suite for measuring latency, peak memory, communication, and numerical equivalence at representative model shapes, by @pds-amd in #14.
-
Preserved Moore Threads MUSA/MCCL support across the expanded distributed VAE and benchmarking infrastructure, by @zoe226 and @pds-amd in #9 and #14.
-
Expanded CPU/Gloo and accelerator-independent testing across supported VAE families, by @pds-amd in #14.
-
Simplified maintenance by wrapping Diffusers modules in place, lazily resolving optional model families, and removing legacy copied implementations, by @pds-amd in #14.
Compatibility
Contributors
Thanks to @TuanNT-ZenAI, @zoe226, @nsakkine, @jjuvonen-amd, and @pds-amd for their contributions to this release.
