Documentation | arXiv | Blog | Benchmark | Gallery

vidax is a lightweight JAX/Flax inference engine and PyTorch-to-JAX weight translator for modern Video Diffusion Transformers (DiTs) and beyond. Built for Google Cloud TPUs, it eliminates framework overhead with clean, explicit PyTree architectures and native multi-chip parallelism (Megatron tensor parallelism and DeepSpeed-Ulysses sequence parallelism) across architecturally distinct model families.

• 🚀 Native TPU performance: jax.sharding device meshes, a real Pallas flash-attention kernel (not jax.nn's materializing default), and a custom scan/vmap-based windowed neighborhood-attention kernel where no native TPU kernel exists.
• 🔄 Universal weight translator: loads PyTorch .safetensors/.pth checkpoints straight into Flax pytrees — key mappings and layout transpositions handled automatically, verified against every model via exact 1:1 parameter-tree matches.
• 🧵 Two parallelism strategies: Megatron-style tensor parallelism and DeepSpeed-Ulysses sequence parallelism, composable and picked per model/resolution depending on whether weight or activation memory is the bottleneck — see docs/hardware_and_sharding.md.
• 💾 Per-layer weight offloading: --offload_dit_weights keeps a DiT's weights host-resident and streams one --offload_chunk_size-block group into HBM at a time, extending every model's reach to resolutions/frame counts that don't fit fully device-resident on a given chip count — see docs/weight_offloading.md.
• 🌊 Flow-matching sampling: deterministic and ancestral (SDE) Euler, plus a from-scratch UniPC multistep predictor-corrector, covering every supported model's native schedule (linear, Karras-sigma, and shift-warped variants alike).
• 🖼️ Faithful conditioning: each model's own image/video-conditioning mechanism ported exactly, not approximated — cross-attention, per-token/ per-frame latent substitution, and conditioning-mask channels all show up where the reference actually uses them.
• 🧩 Broad, growing model coverage: DiTs, dual-pathway Mixture-of-Transformers models, and beyond — every new architecture ported and verified to the same bar (exact checkpoint key/shape matches, bit-exact or real end-to-end checks against the reference).

Current tested TPU types: v4 | v7

Rows are merged across tasks when one script/checkpoint handles all of them (e.g. Wan2.2 TI2V-5B, Cosmos-Predict2.5); kept separate when the reference ships them as genuinely distinct checkpoints/pipelines (e.g. Wan2.1's T2V vs. I2V, Wan2.2's A14B).

Per-model checkpoint sources, CLI flags, and architecture notes live in each Guide link above. Measured latency/memory numbers for every row above live in docs/benchmarking.md.

# Clone and install (editable). On a Cloud TPU VM add the "tpu" extra for the
# right jaxlib wheel: pip install -e ".[tpu]"
git clone https://github.com/FlyingGiraffe/vidax.git
cd vidax
pip install -e .

# Generate a video (Wan2.1 T2V, 1.3B)
python examples/generate_wan2_1_t2v.py \
  --dit_checkpoint_path "./checkpoints/Wan2.1-T2V-1.3B/diffusion_pytorch_model.safetensors" \
  --vae_checkpoint_path "./checkpoints/Wan2.1-T2V-1.3B/Wan2.1_VAE.pth" \
  --t5_checkpoint_path "./checkpoints/Wan2.1-T2V-1.3B/models_t5_umt5-xxl-enc-bf16.pth" \
  --prompt "A majestic red panda climbing a bamboo tree in the snow, 4k" \
  --num_steps 50 \
  --output_path "out/output.mp4"

Beyond the examples/ scripts, vidax is usable as a library — reuse the Pallas flash-attention kernel, the diffusion schedulers, the PyTorch→JAX translator, or a model's DiT/VAE modules directly:

from vidax.core import dot_product_attention, build_tpu_mesh
from vidax.schedulers import RectifiedFlowScheduler
from vidax.translator import load_torch_checkpoint_to_jax

params = load_torch_checkpoint_to_jax("model.safetensors", model_type="wan_dit")
out = dot_product_attention(q, k, v)          # real O(seq)-memory flash attn on TPU

See docs/library_usage.md for worked examples (standalone attention, schedulers, checkpoint translation, and a full model), and docs/api/ for the full per-function API reference.

Standard Python src-layout: one subpackage per model family under models/, one usage guide per family under docs/models/, one standalone inference script per family/task under examples/. See docs/directory_layout.md for the full tree, and docs/index.md for the documentation map.

vidax's source code is released under the Apache License 2.0, except for the JAX re-implementations of HunyuanVideo, HunyuanVideo-1.5 and LTX-2.5. Those were written with reference to upstream code released under the Tencent Hunyuan Community License and the LTX-2.x Community License respectively, and use of those modules is additionally subject to those licenses — which are not permissive open-source licenses (they carry territorial restrictions and acceptable-use / commercial-scale conditions). The affected paths are listed in NOTICE, with a LICENSE file in each affected directory. If you can't accept those terms, don't use those modules; the rest of vidax remains available to you under Apache 2.0.

No model weights are included. Every checkpoint is governed by its own license (the License column below is each model's weights license — the upstream source-code licenses that govern vidax's re-implementations are in NOTICE). You are responsible for complying with the license of any weights you download.

Parallelism techniques implemented in this repo:

• Megatron-style tensor parallelism — Shoeybi et al., Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism.
• DeepSpeed-Ulysses sequence parallelism — Jacobs et al., DeepSpeed Ulysses: System Optimizations for Enabling Training of Extreme Long Sequence Transformer Models.

See docs/hardware_and_sharding.md for how both are implemented here.

This project is supported by the Google TPU Research Cloud (TRC) program.