Skip to content

LingBot-Video Examples

These examples run the Dense 1.3B and MoE 30B LingBot-Video checkpoints for text-to-image, text-to-video, and text-and-image-to-video generation. The MoE path can refine the base result in memory.

Model Source

Model HuggingFace ModelScope Purpose
LingBot-Video Dense 1.3B robbyant/lingbot-video-dense-1.3b Robbyant/lingbot-video-dense-1.3b Dense base generation
LingBot-Video MoE 30B A3B robbyant/lingbot-video-moe-30b-a3b Robbyant/lingbot-video-moe-30b-a3b Mixture-of-experts base generation and refiner

Feature Support

Feature Support Notes
T2I, T2V, and TI2V Supported Select with --task; I2V requires --first_image_path
Multi-GPU inference Supported One or four GPUs with FSDP plus CFG/Ulysses parallelism
MoE refinement Supported MoE video tasks support an in-memory high-resolution refiner
Quantization Partial MoE routed experts support an explicit memory-oriented FP8 backend
CPU offload Supported Used by single-GPU paths and the sequential refiner lifecycle
Feature cache Unsupported No feature cache is configured
Server API Supported Both files declare standard pipeline contracts

Requirements

  • GPU: one CUDA GPU, or four GPUs for the validated distributed layouts
  • Software: the standard TeleFuser installation with compatible Diffusers and Transformers versions
  • Input assets: a structured JSON caption; TI2V also requires a readable first image

Install TeleFuser by following the development setup.

Model Directory

${TF_MODEL_ZOO_PATH}/lingbot/
|-- lingbot-video-dense-1.3b/
|   |-- transformer/
|   |-- processor/
|   |-- text_encoder/
|   |-- vae/
|   \-- scheduler/
\-- lingbot-video-moe-30b-a3b/
    |-- transformer/
    |-- refiner/
    |-- processor/
    |-- text_encoder/
    |-- vae/
    \-- scheduler/

export TF_MODEL_ZOO_PATH=/path/to/model_zoo

Quick Start

python examples/lingbot_video/lingbot_video_dense_1_3b.py \
  --model_root "$TF_MODEL_ZOO_PATH/lingbot/lingbot-video-dense-1.3b" \
  --prompt "$(cat examples/lingbot_video/assets/t2v_5s.json.example)" \
  --task t2v --output_path work_dirs/lingbot-video-dense.mp4

The command writes a five-second, 832x480 landscape video to work_dirs/lingbot-video-dense.mp4.

Examples

Dense Generation

lingbot_video_dense_1_3b.py

python examples/lingbot_video/lingbot_video_dense_1_3b.py \
  --gpu_num 4 --cfg_parallel_degree 2 \
  --model_root "$TF_MODEL_ZOO_PATH/lingbot/lingbot-video-dense-1.3b" \
  --prompt "$(cat examples/lingbot_video/assets/t2v_5s.json.example)" \
  --task t2v --output_path work_dirs/lingbot-video-dense-sp2.mp4

MoE Generation And Refinement

lingbot_video_moe_30b.py

python examples/lingbot_video/lingbot_video_moe_30b.py \
  --gpu_num 4 --cfg_parallel_degree 2 \
  --refiner_gpu_num 4 --refiner_cfg_parallel_degree 2 \
  --refiner_co_resident --expert_backend fp8 --refine \
  --model_root "$TF_MODEL_ZOO_PATH/lingbot/lingbot-video-moe-30b-a3b" \
  --prompt "$(cat examples/lingbot_video/assets/t2v_5s.json.example)" \
  --task t2v --output_path work_dirs/lingbot-video-moe-refined.mp4

CFG parallel and batch CFG are mutually exclusive. Use --no-refiner_co_resident when both DiTs do not fit, or --no-refine for base-only MoE generation. The fp8 expert backend is a memory-oriented option and must be validated on the target GPU.

Serving

telefuser serve examples/lingbot_video/lingbot_video_dense_1_3b.py --port 8000

Serve the MoE model by replacing the script path. TeleFuser creates its workers; do not launch these services with torchrun.