This document is relevant for: Trn2, Trn3

Wan2.2-T2V-A14B Model Card#

Introduction#

Wan2.2-T2V-A14B is a text-to-video diffusion model developed by Wan-AI. It uses a Mixture-of-Experts (MoE) architecture with two experts — a high-noise expert for overall layout during early denoising stages and a low-noise expert for detail refinement during later stages. The model has ~27B total parameters but only 14B active parameters per inference step, keeping computation and memory roughly equivalent to a single 14B dense model. It generates 5-second videos (81 frames at 16fps) at 480P and 720P resolutions with cinematic-level aesthetics and complex motion.

Wan2.2-T2V-A14B is now supported for inference serving with vLLM Omni using the Neuron SDK on AWS Trainium2 (trn2) and Trainium3 (trn3) hardware.

Compatible model checkpoints:

Model	HuggingFace	Hardware	Quantization
Wan2.2-T2V-A14B	Wan-AI/Wan2.2-T2V-A14B-Diffusers	Trn2, Trn3	BF16
Wan2.2-T2V-A14B (FP8 DiT)	Comfy-Org/Wan_2.2_ComfyUI_Repackaged	Trn3	FP8 (`fp8_row_mx`)

FP8 loads a ComfyUI FP8-scaled DiT checkpoint. See Quantization for setup.

Features#

Per-model feature availability for Wan2.2-T2V-A14B. See the README for configuration details.

Category	Feature	Status
Generation	Text-to-Video	✅
	832x480 resolution (480p)	✅
	1280x720 resolution (720p)	✅
	Up to 81 frames	✅
Quantization	BF16	✅
	FP8 (DiT)	✅
Parallelism	Tensor Parallelism (TP)	✅
	Context Parallelism (CP)	✅
	Megatron Sequence Parallelism (SP)	✅
	CFG Parallelism	✅
	VAE Patch Parallelism	✅
Performance	Classifier-Free Guidance	✅
	Spatial Tiling (VAE)	✅
	Temporal Chunking (VAE)	✅
	Continuous request batching	Limited
Compilation	torch.compile	✅

Status legend:

✅ Supported: integrated and tested for Wan2.2-T2V-A14B
Limited: accepted, but concurrent requests run serially rather than as a batched forward pass

Recommended configuration#

The recommended configuration runs on 64 NeuronCores (e.g., trn3), matching the stage config: tensor_parallel_size=4 × ring_degree=8 (context parallelism) × cfg_parallel_size=2 = 64.

CFG parallelism (size=2) runs the conditional and unconditional denoising passes across separate replicas in parallel, reducing wall-clock time per diffusion step.

Context parallelism shards the sequence dimension (temporal-spatial latent tokens) across up to 8 ranks, enabling generation of longer videos within per-device HBM limits. The CP degree is configured with ring_degree, not sequence_parallel_size: vLLM Omni enforces sequence_parallel_size = ulysses_degree * ring_degree, so the stage config sets ring_degree and leaves ulysses_degree (default 1) and sequence_parallel_size unset — vLLM Omni then derives sequence_parallel_size = 1 * ring_degree. ring_degree sizes the sequence-parallel group; the CP self-attention itself runs ring attention (K/V stay sharded and are streamed around the CP group), falling back to all-gather + flash only where the ring kernel cannot run. See the context parallelism design doc for details.

Megatron sequence parallelism (SP) is layered within the TP group (orthogonal to CP): the normalization and MLP regions that TP would otherwise replicate are instead sharded along the sequence dimension across the TP ranks, cutting activation memory at large resolutions. It is enabled with model_config.tp_sequence_parallel: true (see the 720P config wan22_stage_tp4cp8cfg2_720p.yaml) and requires tensor_parallel_size > 1. Leave it unset (the default) at 480P, where activations already fit.

Quantization#

The DiT can run in FP8 to cut weight memory and speed up the projection GEMMs. The mode is fp8_row_mx: it loads a ComfyUI per-tensor FP8-scaled checkpoint and runs the attention (QKV and output) and FFN projections in FP8.

Enable it either from run.py:

python examples/wan22/run.py \
  --quantization fp8_row_mx \
  --comfyui-fp8-model-path Comfy-Org/Wan_2.2_ComfyUI_Repackaged

or in the stage config under engine_args.model_config:

    engine_args:
      model_config:
        quantization: fp8_row_mx
        comfyui_fp8_model_path: Comfy-Org/Wan_2.2_ComfyUI_Repackaged

comfyui_fp8_model_path points at the repo (or local root) holding the ComfyUI high-noise and low-noise DiT checkpoints.

Accuracy Evaluation#

Benchmark: VBench is a comprehensive benchmark suite for video generation models, evaluating across 16 dimensions including subject consistency, motion smoothness, temporal flickering, aesthetic quality, and imaging quality.

See the VBench paper for the benchmark methodology.

Wan2.2-T2V-A14B-Diffusers (BF16)

Subtask and metric	Trn2
Subject Consistency	87.60%
Background Consistency	85.70%
Motion Smoothness	96.50%
Dynamic Degree	61.04%
Appearance Style	27.11%
Scene	47.82%
MLPerf-aligned 6-dimension average	67.63%

For externally published results, see the Wan2.2 VBench evaluation in Compositional Video Generation via Inference-Time Guidance (Table 5) and the VBench Leaderboard. Results from different evaluation configurations are not directly comparable.

Reproduce: Serve the model following the quickstart, then run VBench evaluation:

git clone https://github.com/Vchitect/VBench.git
cd VBench && pip install . && cd ..
python VBench/evaluate.py \
    --videos_path wan_output_dir \
    --dimension subject_consistency background_consistency motion_smoothness \
        dynamic_degree appearance_style scene \
    --mode=custom_input

Known limitations#

Batch size is limited to one. The current Wan2.2 pipeline generates one video per request; concurrent requests run serially rather than as a batched forward pass.
FP8 accuracy validation is in progress. The VBench numbers above are for BF16. FP8 (fp8_row_mx) is functionally supported, but its accuracy has not yet been validated against the BF16 reference.

Tutorials#