wan-pilot — Studio Workbench & Agent Bridge

Presets:

Wan2.2 MoE Engine

75%

Generation Settings

Aspect Ratio

Quality Preset

GPU VRAM Allocation 15.35 GB / 24.00 GB

8.65 GB Free Headroom (Safe) SDPA Native

MoE & Memory Offload FastPath

MoE Sequential Swapping Swaps \(\mathcal{E}_1 \to \mathcal{E}_2\) at \(t_{\text{moe}}\) • Saves ~35GB VRAM

T5 CPU Offload Stores 4.8B text encoder in Host RAM

Tiled 3D Causal VAE Splits spatial decoding into 4 tiles

Model Arch

Resolution

Precision

Latent Dimension: 75,600 tokens

\(T_{\text{latent}} = 21\) • \(H \times W = 45 \times 80\)

Workflow Compiler

Engine Diagnostics Stream

[00:00:01.02] INIT Device: NVIDIA RTX 4090 • 24.00 GB detected

[00:00:02.15] MOE Sequential swap strategy active: \(M_{\text{peak}} \le 18.47\text{ GB}\)

[00:00:03.40] LATENTS Allocated 75,600 tokens across 81 frames

[00:00:04.88] DIFFUSION Step 24/50 • Temporal attention stable