Beidouqixing/minimax-h3-4step-lora-flashgen

MiniMax-H3 4-Step LoRA (FlashGen)

4-step distilled LoRA for MiniMax-H3 text-to-video+audio (T2VA). Trained on Ascend NPU with data-free distillation; merge into the base model and inject base_schedule before vllm-omni / MindIE-SD deployment.

Files

File	Description
minimax_h3_4step_lora_flashgen_v1.0_768p_bf16.safetensors	Student LoRA, rank 64, bf16
merge_lora_ckpt.py	Merge LoRA into base model and inject base_schedule

Training config

Item	Value
Method	Distribution matching distillation (VSD, no GAN), trained on Ascend NPU
Task	T2VA
Inference steps	4
Resolution	1344×768, 5.2s
base_schedule	[1.0, 0.7, 0.4, 0.15, 0.0]
LoRA targets	qkv_proj, out_proj, fc1, fc2, adaln_proj.linear
LoRA rank	64

Usage

Requires base model MiniMaxAI/MiniMax-H3.

Merge with merge_lora_ckpt.py before vllm-omni / MindIE deployment:

python merge_lora_ckpt.py \
  ${MODEL_ROOT}/FL2VA \
  minimax_h3_4step_lora_flashgen_v1.0_768p_bf16.safetensors \
  ./minimax-h3-4step-flashgen

export MODEL=./minimax-h3-4step-flashgen/FL2VA

Deployment (MindIE-SD / vllm-omni)

For Ascend NPU inference and serving, see the MindIE-SD deployment guide:

• MiniMax-H3 推理部署文档（含 4 步蒸馏权重合并与启动）
• MindIE-SD 仓库：Ascend/MindIE-SD