Missing pure Text-to-Video (T2V) inference script for Lyra-2 / Wan2.1-14B on Blackwell (sm_120) via WSL2

AI & Data Science Computer Vision & Image Processing
1

Hi everyone,

I am trying to run pure Text-to-Video generation using the Wan2.1-14B model via the Lyra-2 repository. My environment and hardware are fully prepped and recognized, but I am entirely blocked by the lack of a standard generate.py or Text-to-Video entry point in the Lyra-2 codebase.

I’m hoping someone familiar with this specific fork/repo can point me to the correct launch script.

💻 System & Environment (What Works)

  • Hardware: NVIDIA RTX PRO 6000 (Blackwell, 96GB VRAM).

  • OS: WSL2 (Ubuntu) via Windows.

  • PyTorch: 2.12.0-nightly (Successfully detects the GPU and supports Compute Capability sm_120).

  • Weights: Fully downloaded and verified. The wan2.1-14B checkpoint directory contains all 6 .safetensors shards, the config.json, and the T5 encoder (models_t5_umt5-xxl-enc-bf16.pth).

🛑 The Blocker

We cannot find the actual Python entry point to trigger a simple prompt-to-video generation. The standard scripts found in the original Wan2.1 repository (like generate.py) do not exist in this Lyra-2 folder structure (lyra_2/_src/inference/).

🛠️ Troubleshooting Steps & Trials

We ran traces and executed several scripts in the inference folder, but all either silently failed or were designed for different pipelines:

  1. Tried: lyra_2/_src/inference/lyra2_ar_inference.py

    • Result: Silent exit. We ran a Python trace and confirmed this is just a library file of functions (missing the if __name__ == "__main__": block).

  2. Tried: lyra_2/_src/inference/vipe_da3_gs_recon.py

    • Result: Fails demanding an --input_video_path. This is strictly a 3D Gaussian Splatting video-to-3D reconstruction script, not T2V.

  3. Tried: lyra_2/_src/inference/lyra2_custom_traj_inference.py

    • Result: Fails demanding --input_image_path and --trajectory_path. This is an Image-to-Video camera trajectory pipeline, not pure Text-to-Video.

  4. Searched for T2V Logic:

    • We found wan_t2v_model.py under _src/models/, proving the model class exists in the repo, but we cannot find the corresponding inference script that calls it for a basic text prompt.

The Ask: Has anyone successfully run a pure text prompt to video (no starting image, no input video) using the 14B model within this specific Lyra-2 file structure? If so, what is the exact execution command or script name?

Thanks in advance for the help!

2

hey, with help from awesome **
Gilberto Balderas

I got this running ok.**