Some issues encountered trying to run this on a 5090

[mgrugl]

Hello,

I have been testing LongVie with the Wan2.1-I2V-14B model on an RTX 5090 (WSL2 environment). I encountered several blocking issues that required manual code fixes to get the inference running.

Here is a summary of the bugs and the fixes I applied:

1. Hardcoded Internal Proxy in Shell Script
   In sample_longvideo.sh, the first line exports a private internal proxy (10.66.70.227). Issue: This causes connection timeouts for anyone outside your internal network. Fix: Removed the export http_proxy=... line.

2. Crash Loading Sharded Models (utils.py)
   The Wan2.1-I2V-14B model downloads as multiple .safetensors files (sharded). Issue: load_state_dict in diffsynth/models/utils.py crashes because it expects a single string path, but receives a list of filenames for sharded models. Fix: I patched load_state_dict to handle lists recursively:

# In diffsynth/models/utils.py

def load_state_dict(file_path, torch_dtype=None, device="cpu"):
    # --- ADDED THIS BLOCK ---
    if isinstance(file_path, list):
        state_dict = {}
        for f in file_path:
            state_dict.update(load_state_dict(f, torch_dtype=torch_dtype, device=device))
        return state_dict
    # ------------------------
    
    if file_path.endswith(".safetensors"):
        return load_state_dict_from_safetensors(file_path, torch_dtype=torch_dtype, device=device)
    else:
        return load_state_dict_from_bin(file_path, torch_dtype=torch_dtype, device=device)

3. Typo in T5 Encoder Filename (inference.py)
   Issue: In inference.py, the T5 encoder filename is misspelled as ...models_t5_umt05.... Fix: Changed umt05 to umt5 to match the actual Hugging Face filename (models_t5_umt5-xxl-enc-bf16.pth).

4. Invalid Device Type "gpu" (inference.py)
   Issue: The VAE config in inference.py uses offload_device="gpu". Error: RuntimeError: ... device type at start of device string: gpu Fix: Changed "gpu" to "cuda".

5. Missing Dependency
   Issue: import decord fails. Fix: Added decord to requirements.txt.

6. Folder Naming Inconsistencies
   download_wan2.1.py downloads to a folder named Wan-AI, but inference.py looks for Wan2.1-I2V-14B-480P.

Scripts reference models (plural), but some instructions imply model (singular).

Environment:

GPU: RTX 5090 (requires CUDA 12.8 / Pytorch Nightly cu128 / Flash Attention built from source for sm_120)

OS: WSL2 (Ubuntu 22.04)

Thanks for the great work on the model!

[JianxGao]

Hi @mgrugl,

Thank you for testing LongVie with the Wan2.1-I2V-14B model and for providing such detailed feedback. I really appreciate the time and effort you put into identifying the issues and offering fixes. I apologize for the trouble these issues caused, and I’m grateful for your help in fixing them. I’ll work on addressing these problems as soon as possible to improve the experience for everyone.

Thanks again!

[JianxGao]

I have fixed bugs 1 and 5. However, bugs 2, 4, and 6 may be caused by version updates in DiffSynth-Studio. When we trained the model, we used an older version of the code with DiffSynth version 1.1.8. If you have downloaded the latest version of DiffSynth-Studio, it's possible that these issues were introduced. I will check and fix these as soon as possible(maybe tomorrow). You can refer to this link for the version we used: DiffSynth-Studio v1.1.8

Regarding bug 3, we have not made the error in the code. You can check the reference here: inference.py - Line 30

LongVie/inference.py

Line 30 in 2727f85

ModelConfig(model_id="Wan-AI/Wan2.1-I2V-14B-480P", origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth", offload_device="cpu", skip_download=True),

Thanks again for your thorough testing and valuable feedback. I really appreciate your help!