I’ve been trying to deploy a Python based AWS Lambda that’s using PyTorch. The problem I’ve run into is the size of the deployment package with PyTorch and it’s platform specific dependencies is far beyond the maximum size of a deployable zip that you can deploy as an AWS Lambda. Per the AWS Lambda Limits page, the maximum deployable zip is 50MB (and unzipped it needs to be less than 250MB).
I found this article which suggested to build PyTorch from source in an Amazon AMI EC2, using build options to reduce the build size. I followed all steps up to but not including line 65 as I don’t need torchvision.
If you’re looking for the tl:dr summary, here’s the keypoints:
- yes, this approach works! (although it took many hours spread over a few weeks to get to this point!)
- the specific Amazon AMI you need is the one that’s currently in use for running AWS Lambdas (this will obviously change at some point but as of 9/3/18 this AMI works) : amzn-ami-hvm-2017.03.1.20170812-x86_64-gp2 (ami-aa5ebdd2)
- a t2.micro EC2 instance does not have enough RAM to successfully build PyTorch. I used a t2.medium with 4GB RAM.
- you can’t take a trained model .pt file generated from a different version of PyTorch/torch and use it to generate text using a different version. The PyTorch version for training and generating output must be identical
Ok, beyond the tl;dr summary above, here’s my experience following the steps in this article.
At line 63:
python setup.py install
I got this error:
Could not find /home/ec2-user/pytorch/torch/lib/gloo/CMakeLists.txt
Did you run 'git submodule update --init'?
I ran the suggested ‘git submodule update’ and then re-ran the setup.py script and now it ran for a while but ended with error:
gcc: error trying to exec 'cc1plus': execvp: No such file or directory
error: command 'gcc' failed with exit status 1
I spent a bunch of time trying to work out what was going on here, but I decided to take a different direction and skip building Python 3.6 from source, and try recreating these steps using Python 2.7 that is preinstalled in the Amazon Linux 2 AMI. The only parts that are slightly different is pip is no preinstalled, so I installed it with:
sudo yum install python-pip
sudo yum install python-wheel
The then virtualenv with:
sudo pip install virtualenv
I think point I pick up the steps from creating the virtualenv:
virtualenv ~/shrink_venv
After the step to build pytorch, now I’ve got (another) different error:
as: out of memory allocating 4064 bytes after a total of 45686784 bytes
{standard input}: Assembler messages:
{standard input}:934524: Fatal error: can't close build/temp.linux-x86_64-2.7/torch/csrc/jit/python_ir.o: Memory exhausted
torch/csrc/jit/python_ir.cpp:215:2: fatal error: error writing to -: Broken pipe
Ugh, I’m running in a t2.micro that only has 1GB ram. Let’s stop the instance, change the instance type to a t2.medium with 4GB and let’s try building again.
Running free before:
$ free
total used free shared buff/cache available
Mem: 1009384 40468 870556 288 98360 840700
Swap: 0 0 0
And now after resizing:
$ free
total used free shared buff/cache available
Mem: 4040024 55004 3825940 292 159080 3780552
Swap: 0 0 0
Ok, trying again, but since we’ve rebooted the instance, remembering to set the flags to minimize the build options which was the whole reason we were doing this:
$ export NO_CUDA=1
$ export NO_CUDNN=1
Next error:
error: could not create '/usr/lib64/python2.7/site-packages/torch': Permission denied
Ok, let’s run the build with sudo instead then. That fixes that.
Now I’m at a point where I can actually run the generate.py script but now I’ve got a completely different error:
/home/ec2-user/shrinkenv/lib/python2.7/site-packages/torch/serialization.py:316: SourceChangeWarning: source code of class 'torch.nn.modules.sparse.Embedding' has changed. you can retrieve the original source code by accessing the object's source attribute or set `torch.nn.Module.dump_patches = True` and use the patch tool to revert the changes.
warnings.warn(msg, SourceChangeWarning)
Traceback (most recent call last):
File "generate.py", line 54, in <module>
decoder = torch.load(args.filename)
File "/home/ec2-user/shrinkenv/lib/python2.7/site-packages/torch/serialization.py", line 261, in load
return _load(f, map_location, pickle_module)
File "/home/ec2-user/shrinkenv/lib/python2.7/site-packages/torch/serialization.py", line 409, in _load
result = unpickler.load()
AttributeError: 'module' object has no attribute '_rebuild_tensor_v2'
Searching for the last part of this error found this post, which implies my trained model .pt file is from a different torch/pytorch version … which it most likely is as I trained using a version installed with pip, and now I’m trying to generate with a version built from source.
Rather than spend more time on this (some articles suggested you can read the .pt model from one pytorch version and convert it, but this doesn’t seem like a trivial activity and requires writing some code to do the conversion), so I’m going to train a new model with the same version I just built from source.
Now that’s successfully done, I have my Lambda handler script ready to go, and ready to package up, so back to the final steps from the article to zip up everything built and installed so far in my virtualenv:
cd $VIRTUAL_ENV/lib/python2.7/site-packages
zip -r ~/kevinhookebot-ml-lambda-generate-py.zip *
We’re at 57MB, so looking ok so far (although larger than 50MB?). Now add char-rnn.pytorch, my generated model and Lambda handler into the same zip, and we’re now at 58M so well within the 250MB limit for a Lambda package deployed via S3.
Let’s upload and deploy. Test calling the Lambda, and now we get:
Unable to import module 'generatelambda': /usr/lib64/libstdc++.so.6: version `GLIBCXX_3.4.21' not found (required by /var/task/torch/lib/libshm.so)
Searching for this error I found this post which has a link to this page which lists a specific AMI version to be used when compiling dependencies for a Lambda deployment (amzn-ami-hvm-2017.03.1.20170812-x86_64-gp2). Picking the default Amazon Linux 2 AMI is probably not this same image (and I already tried the Amazon Linux AMI 2018.03.0 and ran into other issues on that one) so looks like I need to start over (but getting close now, surely!)
Ok, new EC2 t2-medium instance with the exactly the same AMI image as mentioned above. Retraced my steps and now I feel I’m almost back at the same error as before:
error: command 'gcc' failed with exit status 1
gcc: error trying to exec 'cc1plus': execvp: No such file or directory
Searching some more for this I found this post with a solution to change the PATH to point exactly where cc1plus is installed. Instead of 4.8.3 in this AMI though it seems I have 4.8.5, so here’s the settings I used:
$ export COMPILER_PATH="/usr/libexec/gcc/x86_64-amazon-linux/4.8.5/:$COMPILER_PATH";
export C_INCLUDE_PATH="/usr/lib/gcc/x86_64-amazon-linux/4.8.5/include/:$C_INCLUDE_PATH";
And then I noticed in the post they hadn’t included either of these in setting the new PATH which seems like an oversight (I don’t think these will make any difference if they are not in the PATH), so I set my path like this, including COMPILER_PATH first:
export PATH="$COMPILER_PATH:/sbin:/bin:/usr/sbin:/usr/bin:/opt/aws/bin:/usr/local/bin:/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/sbin:/opt/aws/bin:$PATH";
Now cc1plus is in my path, let’s try building pytorch again! In short, that worked.
Going back to the packaging steps, let’s zip everything up and push the zip to S3 ready to deploy.
I zipped up:
- char-rnn.pytorch : from github, including my own runner script and lambda handler script
- modules installed into virtualenv lib: ~/shrinkenv/lib/python2.7/site-packages/* (tqdm)
- modules installed into virtualenv lib64: ~/shrinkenv/lib64/python2.7/site-packages/* (torch, numpy and unidecode)
At this point I used ‘aws s3 cp’ to copy the zip to s3, and configured my Lambda from the zip in s3. Set up a test to call my handler, and success!