Basic Inference Quickstart¶
This guide validates a maintained single-GPU text-to-video example and the batch HTTP service. For TeleFuser's primary interactive workflow, start with the LingBot-World v2 WebRTC Core Experience.
Prerequisites¶
- A Linux checkout of the TeleFuser repository
- Python 3.10 through 3.13 and a working CUDA-enabled PyTorch installation
- One CUDA GPU with enough memory for Wan2.1 T2V 1.3B at 480p
- An existing local Wan2.1 T2V 1.3B checkpoint; this workflow does not download models
Follow Installation first. Confirm that torch.cuda.is_available() returns True.
Run the Pipeline¶
From the repository root:
mkdir -p work_dirs
export WAN21_MODEL_SOURCE=/path/to/model_zoo/Wan2.1-T2V-1.3B
TELEAI_EXAMPLE_OUTPUT_DIR=work_dirs \
python examples/wan_video/wan21_1_3b_text_to_video_hf.py \
--model_root "$WAN21_MODEL_SOURCE" \
--resolution 480p \
--prompt "A sailboat crosses a calm lake at sunrise"
A successful run ends with a Video saved to: message and writes:
The checkpoint is also published as Wan-AI/Wan2.1-T2V-1.3B on Hugging Face and Wan-AI/Wan2.1-T2V-1.3B on ModelScope. Set WAN21_MODEL_SOURCE to the existing local repository directory without flattening its layout. The links above identify the checkpoint; they are not download steps in this guide.
Start the Batch Service¶
In the shell where WAN21_MODEL_SOURCE is set, keep this process running:
Wait for startup to complete, then check the service from another terminal:
Create a task:
curl --fail --request POST http://127.0.0.1:8000/v1/tasks/create \
--header "Content-Type: application/json" \
--data '{
"task": "t2v",
"prompt": "A sailboat crosses a calm lake at sunrise",
"resolution": "480p",
"aspect_ratio": "16:9"
}'
The response contains a task_id, task_status, and output_path. Poll the returned task ID until status is completed:
The running server also publishes Swagger UI at http://127.0.0.1:8000/docs and its OpenAPI document at http://127.0.0.1:8000/openapi.json.
Next Steps¶
- Supported Models lists model families and task-specific guides.
- Serving and APIs explains batch and streaming modes.
- Runtime and optimization covers configuration, parallelism, attention, and caching.
- Troubleshooting provides symptom-based diagnostics.
Continue with the LingBot-World v2 WebRTC Core Experience to exercise the framework's stateful, bidirectional streaming path.