What Are the Required Library Versions (diffusers/transformers/peft/torch) for Stable Diffusion 3.5?

[gWeiXP]

Could you please advise on the specific environment configuration required for using Stable Diffusion 3.5 (especially the stable-diffusion-3.5-large-turbo model)? In particular, what are the compatible version combinations of core libraries such as diffusers, transformers, peft, and torch?

The requirements.txt file in the official GitHub code does not specify exact versions. However, updating all libraries to their latest versions results in various errors stating that certain modules cannot be imported from specific libraries.

Additionally, both the official code and online tutorials only mention updating diffusers to the latest version. Currently, my torch version is 2.2.0, and the latest compatible diffusers version I can install is 0.35. Do I need to update my torch version? From what I can see, the weight files on Hugging Face are all from last year.

I managed to successfully import [the necessary modules] using an older version combination, but then encountered an error indicating a mismatch between the imported model and the weight files—this is presumably also a version-related issue.

Therefore, I would like to know the specific versions of libraries (diffusers, transformers, peft, torch) required for Stable Diffusion 3.5.

[ricber]

I was able to successfully run the following command:

python3 sd3_infer.py --prompt "cute wallpaper art of a cat"

Environment

• Conda: 23.3.1
• CUDA: 11.8.0
• GCC: 14.2.0
• Python: 3.10

Pip packages

Here’s an export of the pip-installed packages in this environment:

einops==0.8.1
fire==0.7.1
numpy==2.2.2
pillow==11.0.0
protobuf==5.29.3
sentencepiece==0.2.0
tiktoken==0.11.0
tokenizers==0.22.0
torch==2.5.1+cu118
torchvision==0.20.1+cu118
transformers==4.56.1
webdataset==1.0.2

---

Note

You may need to update the tokenizer filename in the sd3_infer.py file.

Change this line:

f"{model_folder}/t5xxl.safetensors", framework="pt", device="cpu"

to the actual filename, likely:

f"{model_folder}/t5xxl_fp16.safetensors", framework="pt", device="cpu"