Skip to content

README.md

Latest commit

 

History

History
31 lines (20 loc) · 1.86 KB

README.md

File metadata and controls

31 lines (20 loc) · 1.86 KB
Raw

Self-Forcing (Wan2.1-T2V-1.3B)

This page shows how to run the Self-Forcing Wan2.1 1.3B text-to-video model in LightX2V.

Prepare the model

Follow the LightX2V Quick Start. Download the original Wan2.1 1.3B weights and the Self-Forcing checkpoint:

hf download Wan-AI/Wan2.1-T2V-1.3B --local-dir ./wan_models/Wan2.1-T2V-1.3B
hf download gdhe17/Self-Forcing checkpoints/self_forcing_dmd.pt --local-dir .

The current RTX 5090 profiles use the same 81-frame, 480×832, 16 FPS request geometry and LightTAE decoder:

The FP8 configuration uses converted DiT and T5 checkpoints. Adjust the absolute weight, tokenizer and decoder paths in a copied JSON before running it on another machine. Both profiles enable Sage, warmup, DiT compile/CUDA Graph, streaming LightTAE and top-level asynchronous decode. KV offload is disabled; vae_use_compile controls the decoder independently of use_compile.

Run

Run from the repository root and point MODEL_PATH at the downloaded Wan weights. The launchers find the repository automatically and select the respective configurations above:

export MODEL_PATH="$PWD/wan_models/Wan2.1-T2V-1.3B"
bash scripts/self_forcing/run_wan_t2v_sf.sh
bash scripts/self_forcing/run_wan_t2v_sf_fp8.sh

To select a copied JSON, set CONFIG_PATH when invoking either launcher. The source checkpoint path is the dit_original_ckpt field. Self-attention writes RoPE-rotated keys to a rolling ring; Sage stores encoded historical K/V and incrementally updates the newly written slots. See the AR KV implementation map for the storage and FlashAttention paths.