Wan-AI

/

Wan2.1-T2V-14B

Like 1.62k

Follow

Wan-AI 15k

Hi everyone,

For creators and developers who want to run Wan2.1 14B locally without needing 48GB/80GB enterprise GPUs, we built and open-sourced PhobosStream:

🔗 GitHub Repository: https://github.com/martianacademy/phobos-stream
🎨 ComfyUI Node: https://github.com/martianacademy/ComfyUI-PhobosStream

Key Capabilities:

• Zero-Copy mmap DiT Streaming: Instead of allocating 28–38 GB persistent VRAM for the full DiT, weights remain memory-mapped on NVMe SSD and stream block-by-block during the denoising loop.
• Sequential Stage Offloading: T5-XXL text encoder executes in Stage 1 and is immediately evicted before DiT denoising starts.
• Live Empirical Telemetry: On Apple Silicon Metal GPU (720p HD latents), peak memory is 788.11 MB (>97% reduction from baseline).
• Universal Hardware: Works across Apple Silicon (MLX), NVIDIA CUDA, and ComfyUI.

Hope this helps community members run Wan2.1 14B on 8GB/12GB consumer GPUs and 16GB/24GB Macs!