I2V model running in fp16 produces only noise

[realisticdreamer114514]

System Info / 系統信息

diffusers 0.30.3 on cuda 12.4 & python 3.11 in conda venv of windows

Information / 问题信息

• [X] The official example scripts / 官方的示例脚本
• [ ] My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

1. In cli_demo.py, add these optimizations

   pipe.enable_model_cpu_offload()
   pipe.enable_sequential_cpu_offload()
   pipe.vae.enable_slicing()
   pipe.vae.enable_tiling()

   and reduce the number of frames to 41 (for 5 seconds)

2. Run cli_demo.py python cli_demo.py --prompt "A female basketball player is standing in a basketball court, her body leaning forward towards the camera. She is wearing a vibrant blue basketball jersey with the number 3 prominently displayed. The person's head is tilted back, and her hands are clasped together in front of her legs. She is reaching for the viewpoint and waving at it closely. The court beneath her is a rich brown color with green wall in the background and a basketball hoop stands against the wall." --model_path "D:\CogVideoX-5b-I2V" --generate_type "i2v" --output_path ./output.mp4 --image_or_video_path "D:\test\process\2.png" --dtype float16
3. Video has only colored noise

Expected behavior / 期待表现

A normal video is produced

[kijai]

The I2V model will generally do exactly that if anything but 49 frames is used.

[realisticdreamer114514]

The I2V model will generally do exactly that if anything but 49 frames is used.

This is the case, thanks for pointing out.