What version of pytorch should be?

[VitaliyAT]

4090
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
and
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
gave the same result

UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, �[1mdo those steps only if you trust the source of the checkpoint�[0m. (1) In PyTorch 2.6, we changed the default value of the weights_onlyargument intorch.loadfromFalsetoTrue. Re-running torch.loadwithweights_onlyset toFalsewill likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source. (2) Alternatively, to load withweights_only=Trueplease check the recommended steps in the following error message. WeightsUnpickler error: Unsupported global: GLOBAL functools.partial was not an allowed global by default. Please usetorch.serialization.add_safe_globals([partial])or thetorch.serialization.safe_globals([partial]) context manager to allowlist this global if you trust this class/function. Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.

It ask me do some changes somewhere I don;t understand where. Where is torch.load placed? Why not do something by default?

[DCY1117]

The error message says that the newest pytorch version 2.6.0 changed the default value of weights_only in torch.load() to True, which is likely causing compatibility issues. Since I'm using the FastAI library to load models, it may not have accounted for this change yet. The quickest fix would be to downgrade to an older PyTorch version, such as 2.5.0, to avoid this issue.

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
and
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

you should pick the ones that matches your cuda version

As for the old pytorch version check the official page:
https://pytorch.org/get-started/previous-versions/

[NSpecies]

I'm also having this issue, and tried all of the recommended changes bringing pytorch to 2.5.0 and trying both pip3 installs. I don't suppose there is a way to pass an argument through console to set weights_only? Is there a good way to convert the pkl files for the text detection in to a format that doesn't cause torch to get angry? I've been trying but nothing works for me. Maybe you can do it since this'll run.

[DCY1117]

Make sure you have installed an older version of PyTorch. PyTorch 2.5.0 works fine for me, and you should only see the following warning:

FutureWarning: You are using torch.load with weights_only=False (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for weights_only will be flipped to True. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via torch.serialization.add_safe_globals. We recommend you start setting weights_only=True for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
state = torch.load(source, map_location='cpu') if defaults.device == torch.device('cpu') else torch.load(source)

If you still encounter the same error message, it means you are still using PyTorch 2.6.0. The easiest fix right now is to downgrade to an older version, which is just a simple command to install an earlier release. Follow the official instructions here: https://pytorch.org/get-started/previous-versions/

I've also updated the README to reflect this issue.

[H644b]

I still got this error:

ValueError: Due to a serious vulnerability issue in torch.load, even with weights_only=True, we now require users to upgrade torch to at least v2.6 in order to use the function. This version restriction does not apply when loading files with safetensors. See the vulnerability report here https://nvd.nist.gov/vuln/detail/CVE-2025-32434

[H644b]

Was able to bypass this by adding this to import_utils.py in the "transformers" package:

def check_torch_load_is_safe(): return

[h4rm0n1c]

okay, fair warning, this really sucks, and I suck for doing this, as of now I've had to resort to this ugliness:
add these lines around line 33 of MangaQuick.py:

import torch
from pathlib import Path
from fastai.basic_train import load_learner as fastai_load_learner
from fastai.torch_core import defaults

def patched_load_learner(path, file, device=None):
    path = Path(path)
    file = Path(file)
    if device is None:
        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    original_torch_load = torch.load

    def patched_torch_load(*args, **kwargs):
        kwargs['weights_only'] = False
        return original_torch_load(*args, **kwargs)

    torch.load = patched_torch_load
    try:
        learner = fastai_load_learner(path, file, device=device)
    finally:
        torch.load = original_torch_load

    return learner

# Override load_learner globally
load_learner = patched_load_learner

modify /components/image_inpainting/bin/saicinpainting/training/trainers/init.py:
replace load_checkpoint with:

from torch.serialization import safe_globals
from pytorch_lightning.callbacks.model_checkpoint import ModelCheckpoint

def load_checkpoint(train_config, path, map_location='cuda', strict=True):
    model: torch.nn.Module = make_training_model(train_config)
    
    # Allowlist ModelCheckpoint for safe unpickling
    with safe_globals([ModelCheckpoint]):
        state = torch.load(path, map_location=map_location, weights_only=False)

    model.load_state_dict(state['state_dict'], strict=strict)
    model.on_load_checkpoint(state)
    return model

Really ugly, really bad, but it works, absolutely hacky, dangerous, even, you are going to load the entire model, code and all with this.
I do not recommend exposing this to the internet, in this case you'll want to run your app with
streamlit run MangaQuick.py --server.address 0.0.0.0
to force local LAN operation only.
I'll make this clear, this gets it up and running but this is dangerous, it's insecure, you can get hacked by a malicious model by doing this.
This repo needs a security patch for pytorch 2.6.0 but these kinds of monkey patches are about the limit of my capabilities.