Feature Request: Support Multi-GPU Parallelism (TP) for Faster Video Inference

Hi @lllyasviel,

First, thank you for the amazing work on FramePack! It’s impressive how it enables low-VRAM inference for AI video generation.

I’d like to request support for ​multi-GPU parallelism​ (e.g., Tensor Parallelism) to further reduce per-video inference time, especially for longer/higher-resolution videos. Here’s why this would be valuable:

​Speedup: Distributing workloads across GPUs (e.g., via TP for model sharding or DP for batch processing) could significantly cut latency.
​Scalability: Users with multi-GPU setups (e.g., 2x/4x RTX 4090) could leverage unused resources.

​Suggested Approaches:
​Tensor Parallelism (TP)​: Split model layers (e.g., attention heads, MLPs) across GPUs. For example, 4-way TP could shard QKV projections column-wise.

​Questions:
- Is multi-GPU support already planned? If so, what’s the roadmap?
- Would you accept PRs implementing this (e.g., via PyTorch’s DistributedDataParallel or Megatron-style TP)?

Thanks for considering! This could make FramePack even more powerful for production use.

The issue names no files, tests, or entry points, so begin by locating the video inference and model-loading paths. Clarify whether tensor parallelism, data parallelism, or both are required, then define supported multi-GPU configurations and verification for reduced inference time.

Written by the indexing model from the issue text.