OSError: Consistency check failed

[Nirbhay2727]

Describe the bug

OSError: Consistency check failed: file should be of size 9860464979 but has size 4277965032 ((…)l-00006-of-00007.bin). We are sorry for the inconvenience. Please retry download and pass force_download=True, resume_download=False as argument. If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub. Downloading (…)l-00006-of-00007.bin: 43%|███████████████▌ | 4.28G/9.86G [03:25<04:27, 20.8MB/s]

Reproduction

No response

Logs

model = AutoModelForCausalLM.from_pretrained(checkpoint,force_download=True, resume_download=False).to(device)
#device='cuda'

System info

- huggingface_hub version: 0.15.1
- Platform: Linux-5.15.0-1033-oracle-x86_64-with-glibc2.35
- Python version: 3.10.11
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Token path ?: /home/ubuntu/.cache/huggingface/token
- Has saved token ?: True
- Who am I ?: Nirbhay2727
- Configured git credential helpers: 
- FastAI: N/A
- Tensorflow: N/A
- Torch: 1.13.1
- Jinja2: N/A
- Graphviz: N/A
- Pydot: N/A
- Pillow: 9.4.0
- hf_transfer: N/A
- gradio: N/A
- numpy: 1.24.3
- ENDPOINT: https://huggingface.co
- HUGGINGFACE_HUB_CACHE: /home/ubuntu/.cache/huggingface/hub
- HUGGINGFACE_ASSETS_CACHE: /home/ubuntu/.cache/huggingface/assets
- HF_TOKEN_PATH: /home/ubuntu/.cache/huggingface/token
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False

[Wauplin]

Hi @Nirbhay2727, I'm sorry for the inconvenience. Can you tell me which model and file you are trying to instantiate ? Thanks in advance

[mateoriosglb]

Hello, I'm having the same issue with models:

• stabilityai/stable-diffusion-2-1
• runwayml/stable-diffusion-v1-5
• SG161222/Realistic_Vision_V1.4
• prompthero/openjourney

[ParseDark]

same issue. I want to use the whisper

[matthew-hippocraticai]

Same issue with falcon-40b-instruct

[Wauplin]

Sorry not taking care of it earlier. @matthew-hippocraticai could you post here:

• the output of huggingface-cli env (run in your machine)
• the script you used to download the weight
• the exact stacktrace you get (the error message + trace)

If you can't reproduce the error please let me know as well. And just in case, could you update to latest huggingface_hub version and retry?

Thanks a lot in advance! I'll investigate this thoroughly in the coming days so more information will help.

[matthew-hippocraticai]

I believe the issue was actually just running out of storage. Has been fixed now, thanks!

[Wauplin]

I believe the issue was actually just running out of storage.

Ah! That's a really good insight for us. How did you found out in the end? @mateoriosglb @ParseDark could it be the same for you?

If confirmed, I'll add a message to the exception to suggest to the user to check disk usage and retry.

[matthew-hippocraticai]

It was simply OSError: [Errno 28] No space left on device but I didn't see it at first and just assumed this error was the only one

[julien-c]

note that we could maybe explore checking space left on device before starting downloading (but – and it's a big but – i'm not sure if there's any reliable way to do this)

[Wauplin]

@julien-c Given https://github.com/huggingface/huggingface_hub/issues/1498#issuecomment-1623926541, I don't think space left on device was the root cause of the problem for everyone in this issue (ping @mateoriosglb @ParseDark @Nirbhay2727 if you can confirm?). But anyway, checking for free space is still an interesting nice to have IMO so I opened an issue for it https://github.com/huggingface/huggingface_hub/issues/1551. There is a cross-platform Python's built-in to check disk_usage. I would only trigger a warning instead of raising an exception.

[Wauplin]

Also related, I've got a user getting repeatedly this issue (https://github.com/huggingface/huggingface_hub/issues/1549) but not when downloading with snapshot_download. I'm trying to investigate it more.

[mateoriosglb]

I still don't know the source of the problem but I agree with @Wauplin. I have more than enough space left so I rebooted the kernel, reloaded everything and then they started working fine. However I would like to find the source because it is possible that it will happen again.

[978551022]

I've been trying to re-download it for 5 days, and it always runs out I changed everything I could and I still solved it like this

[978551022]

[Nirbhay2727]

The issue was solved for me by downgrading bitsandbytes to 0.37.0. I believe that the cause of the issue was that a deep copy of the model was created while saving the checkpoints.

[Wauplin]

@978551022 I'm sorry you are having troubles here but if I understand correctly, this is not a problem of consistency issue, right? If you are download files with a slow connection, I would highly recommend you to use resume_download=True. If some timeout errors happen, you "just" have to relaunch it. This can even be done in a single script:

import requests
from huggingface_hub import snapshot_download

while True:
    try:
        snapshot_download(..., resume_download=True)
        break
    except requests.ReadTimeout as e:
        print(f"Read timeout: {e}")

Each time the snapshot_download is restarted, it takes a bit of time to list the files to download but this init phase should be neglectable compared to the download time.

I'm sorry that there's no built-in solution right now but error handling is not always easy to integrate to a library - as user expectations might differ. Hope this snippet of code will make your life easier :)

[978551022]

@978551022很抱歉您在这里遇到了麻烦，但如果我理解正确的话，这不是一致性问题，对吧？如果您要使用较慢的连接下载文件，我强烈建议您使用resume_download=True. 如果发生一些超时错误，您“只需”重新启动它。这甚至可以在单个脚本中完成：

import requests
from huggingface_hub import snapshot_download

while True:
    try:
        snapshot_download(..., resume_download=True)
        break
    except requests.ReadTimeout as e:
        print(f"Read timeout: {e}")

每次重新启动 snapshot_download 时，都会花费一些时间来列出要下载的文件，但与下载时间相比，此初始化阶段应该可以忽略不计。

很抱歉，目前没有内置的解决方案，但错误处理并不总是很容易集成到库中 - 因为用户的期望可能有所不同。希望这段代码能让您的生活更轻松:)

Breakpoint reconnection This configuration I have set to True But still timeout, I even changed socket and requests, urllib3 source code Still to no avail In theory, resume_download=True can solve this problem But thanks for your reply. I'm using it now dataset = load_dataset("mlfoundations/datacomp_pools",num_proc=64)

It is currently being downloaded The one I used before won't work: nohup python download_upstream.py --scale xlarge --data_dir /home/nfs-s2/datacomp_pools/xlarge/ --skip_shards --download_npz --processes_count 12768 --thread_count 100000 > /home/wx1271473/code/download.log 2>&1 &

[Wauplin]

Then maybe the best is to do the retry loop around load_dataset. At least you wouldn't have to manually restart the script each time. Apart from that, I don't really have a solution to offer

while True:
    try:
        dataset = load_dataset("mlfoundations/datacomp_pools",num_proc=64)
        break
    except Exception as e:
        print(f"Exception: {e}")

[978551022]

现在也许最好的方法是进行重新试循环load_dataset。至少您不必每次都手动重新启动脚本。除此之外，我真的没有提供解决方案

while True:
    try:
        dataset = load_dataset("mlfoundations/datacomp_pools",num_proc=64)
        break
    except Exception as e:
        print(f"Exception: {e}")

Okay, this is a violent download So far I've been able to run it stably load_dataset("mlfoundations/datacomp_pools",num_proc=64)

[kanav1504]

Hi, I have the same error trying to download mistralai/Mixtral-8x7B-v0.1.

OSError: Consistency check failed: file should be of size 4983004072 but has size 4124798953 (model-00011-of-00019.safetensors). We are sorry for the inconvenience. Please retry download and pass force_download=True, resume_download=False as argument. If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.

[brunopistone]

Hello, I want to report a similar issue with mistralai/Mixtral-8x7B-v0.1.. I've opened a discussion also here

bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16
    )

    model = AutoModelForCausalLM.from_pretrained(
       "mistralai/Mixtral-8x7B-Instruct-v0.1",
        trust_remote_code=True,
        force_download=True,
        resume_download=False,
        quantization_config=bnb_config,
        device_map="auto")

Error: AlgorithmError: OSError('Consistency check failed: file should be of size 4221679088 but has size 3663094841 (model-00019-of-00019.safetensors).\nWe are sorry for the inconvenience. Please retry download and pass force_download=True, resume_download=False as argument.\nIf the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.'), exit code: 1

I've noticed that when the shards are loaded, it appears the following warning UserWarning: Not enough free disk space to download the file. The expected file size is: 4221.68 MB. The target location /root/.cache/huggingface/hub only has 3747.39 MB free disk space.. I have more than 400 GB of free space, so it's not even possible that there is missing space on the device.

Python modules: transformers==4.37.2 peft==0.7.1 accelerate==0.26.1 bitsandbytes==0.42.0

[Wauplin]

@brunopistone I'm sorry you're facing this issue. I would be surprised if the UserWarning: Not enough free disk space to download the file. The expected file size is: 4221.68 MB. The target location /root/.cache/huggingface/hub only has 3747.39 MB free disk space. warning is raised but that you have enough free space on that drive. Are you sure the 400GB are on /root/.cache. What is the output of df -h /root/.cache for you?

If the cache directory is on a small partition, you can always set the cache to a different volume with more disk space available by setting the HF_HOME environment variable (see documentation).

[haochengxi]

If you use ln -s /.cache /root/,cache since the space of /root/ is limited, you need to export HF_HOME=. This solves my problem.

[JohnHerry]

Hi, I have the same error trying to download mistralai/Mixtral-8x7B-v0.1.

OSError: Consistency check failed: file should be of size 4983004072 but has size 4124798953 (model-00011-of-00019.safetensors). We are sorry for the inconvenience. Please retry download and pass force_download=True, resume_download=False as argument. If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.

The same error when downloading a very large dataset. Can we make the Consistence check Error a Consistence check Warning so that other part of the dataset can be downloaded continully?

[Qingrenn]

Hi, I have the same error trying to download mistralai/Mixtral-8x7B-v0.1. OSError: Consistency check failed: file should be of size 4983004072 but has size 4124798953 (model-00011-of-00019.safetensors). We are sorry for the inconvenience. Please retry download and pass force_download=True, resume_download=False as argument. If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.

The same error when downloading a very large dataset. Can we make the Consistence check Error a Consistence check Warning so that other part of the dataset can be downloaded continully?

Yes, I also met the same error when downloading the large dataset Salesforce/lotsa_data. Can we redownload the corrupted files after other part of the dataset have been downloaded ?

Exception: Consistency check failed: file should be of size 488482328 but has size 55485239 ((…)vm_traces_2017/data-00004-of-00022.arrow).
We are sorry for the inconvenience. Please retry download and pass `force_download=True, resume_download=False` as argument.
If the issue persists, please let us know by opening an issue on https://github.com/huggingface/huggingface_hub.

[fblgit]

same problem here, weird stuff.. here is here any dos2unix issue kind of thing? first time i see this issue..

[Wauplin]

Can we redownload the corrupted files after other part of the dataset have been downloaded ?

Downloading files using datasets is slightly different at the moment (not the same cache). Ping @lhoestq is it possible to force trigger the redownload of a single arrow file?

[Wauplin]

first time i see this issue..

@fblgit I'm sorry for the inconvenience. It can be related to network issues but hard to investigate. Best strategy is still to redownload the corrupted file. Not sure about the dos2unix TBH. It could be possible for small text files but not huge binaries (which is where is usually happen).

[lhoestq]

you can force-redownload the file manually, e.g. in your case

from datasets import DownloadConfig, DownloadManager

url = "hf://datasets/Salesforce/lotsa_data/azure_vm_traces_2017/data-00004-of-00022.arrow"
config = DownloadConfig(force_download=True)
DownloadManager(config).download(url)

[nguyenkhangme]

I got the same problem and in my case, it turns out the downloading process has been terminated so when it starts again, it just adds up to the same file, which makes "Consistency check failed". Solved this by:

pip install huggingface_hub\[cli\]
huggingface-cli delete-cache

then delete the cache and download again