Perflow-Shuai/LongLive-LoRA-nonAR-DMD-RGS3-iter1500

LongLive-LoRA non-AR DMD — RGS 3 · Iter 1500

Generator LoRA for four-step, full-sequence non-autoregressive inference with Wan-AI/Wan2.2-TI2V-5B. This release is the literal training checkpoint at iteration 1500 from the 16-GPU dev/nonAR experiment with real_guidance_scale=3.

real_guidance_scale is a training setting. The distilled generator was validated with No CFG at inference (guidance_scale=1, one conditional model call per denoising step).

Checkpoint identity

Item	Value
Base model	Wan-AI/Wan2.2-TI2V-5B
Architecture	full-sequence non-AR
Objective	DMD with backward simulation
Training RGS	3
Literal checkpoint	checkpoint_model_001500
Training world size	16
LoRA rank / alpha	128 / 128
Generator LoRA	600 FP32 tensors, 322,437,120 parameters
Inference	4 denoising steps, No CFG

Training used the vidprom_filtered_extended.txt prompt source and saved/evaluated every 250 iterations.

Files and usage

• generator_lora.pt: native LongLive wrapper with generator_lora and step.
• adapter_model.safetensors: the same 600 generator tensors in safe format.
• adapter_config.json: PEFT rank/alpha and all 300 target modules.
• inference_config.yaml: essential non-AR, four-step, No-CFG overrides.
• training_config.yaml: the recorded training configuration.
• provenance.json: source and published-file checksums.

from huggingface_hub import hf_hub_download

lora_path = hf_hub_download(
    repo_id="Perflow-Shuai/LongLive-LoRA-nonAR-DMD-RGS3-iter1500",
    filename="generator_lora.pt",
)

Use lora_path as checkpoints.lora_ckpt in LongLive-LoRA. Keep generator_is_causal: false, LoRA rank/alpha 128/128, sampling steps 4, and inference guidance scale 1.0. Do not use the generic causal defaults for this adapter.

Provenance note

The published files are a lossless generator-only extraction. The original exact-resume checkpoint also contains a training-only critic LoRA; optimizer and per-rank RNG/dataloader state are intentionally not published here.

Source checkpoint SHA-256: a79c44a4039846d005fcc71285b4d88117957b503b0b9332d354f0a2e8acfc7c. Training provenance records Git commit 655718bd6b27950b20c269b2c3275d786cc75348 together with source-tree digest 39887517aa21ab0e6b14021c1cd6ff13f293adb39ae07c9c7fd28dd89a324d0a; the commit alone is not claimed to reconstruct the complete training snapshot.