OOM kill on pip install

[ezyang]

When I attempt to install Torch with pip, the process gets OOM killed:

ezyang@sabre:~/Dev$ pip install https://download.pytorch.org/whl/cu75/torch-0.1.10.post2-cp27-none-linux_x86_64.whl 
Collecting torch==0.1.10.post2 from https://download.pytorch.org/whl/cu75/torch-0.1.10.post2-cp27-none-linux_x86_64.whl
  Downloading https://download.pytorch.org/whl/cu75/torch-0.1.10.post2-cp27-none-linux_x86_64.whl (360.3MB)
    99% |████████████████████████████████| 360.3MB 32.1MB/s eta 0:00:01Killed

dmesg:

[326093.958653] Out of memory: Kill process 14093 (pip) score 452 or sacrifice child
[326093.958669] Killed process 14093 (pip) total-vm:5029612kB, anon-rss:4217296kB, file-rss:4kB

Feel free to close this if it is by design that 4G is not enough memory to install Torch (yes I don't have very much memory) but perhaps there is something here worth investigating?

[soumith]

thanks for reporting this.

We ship binaries, which means that all pip is doing is: unzipping the package on your computer and moving files into the right location.

With this context, there is very little I see I can do from the package side to improve this.

When I get time, I can try to simulate this in a limited memory environment and try to find out the exact reasons and report something to the pip folks upstream, but as the task is not really actionable from my side, I'll close this issue.

[diwu1989]

You can wget the whl package locally and then run pip install on it, that seems to use less memory

[rkingery]

I'm still seeing this problem when trying to pip install torch inside the base Ubuntu Docker container. It'll get 99% installed and then kill. Other packages install fine.

[vikramriyer]

I am facing the exact issue as one faced by @rkingery.

[heiner]

Same issue here.

Can be fixed by increasing Docker memory (e.g. https://stackoverflow.com/questions/44533319/how-to-assign-more-memory-to-docker-container).

@diwu1989's comment is interesting though: Is there a less demanding way of downloading and installing PyTorch than pip?

[heiner]

(Turns out a pip download torch followed by pip install torch*.whl does not go OOM for me.)

[wpietri]

In case others need a workaround, pip townload torch OOMed for me. Instead I had to:

• Google the thing it was downloading
• Visit that page on Wheelodex; for me that was https://www.wheelodex.org/projects/torch/wheels/torch-1.6.0-cp38-cp38-manylinux1_x86_64.whl/
• grab the download link; for me that was https://files.pythonhosted.org/packages/8c/5d/faf0d8ac260c7f1eda7d063001c137da5223be1c137658384d2d45dcd0d5/torch-1.6.0-cp38-cp38-manylinux1_x86_64.whl
• manually download that with wget
• do pip install torch-1.6.0-cp38-cp38-manylinux1_x86_64.whl

This suggests to me that the problem here is a pip bug; it must be allocating a lot of memory when it apparently doesn't need to.

[perber]

I got the same issue. Another way to install the package is to use the --no-cache-dir option.
It worked on our environment.
pip --no-cache-dir install torch
Hopefully this helps some one.

[Horsmann]

Same problem @perber's solution worked for me in my docker container

[espears1]

I am experiencing this same problem but --no-cache-dir does not solve it.

[bthiban]

RUN pip install -r requirements.txt --no-cache-dir when your torch is inside requirements.txt

[espears1]

@bthiban I mentioned that in #1022 (comment) but unfortunately is did not solve the problem.