Installation of the torch package to isolated python environment about neuropod HOT 2 OPEN

Including a requirements.lock file could resolve the issue but this would cause installation when loading the model.

As you noted, passing in requirements to create_python_neuropod or create_pytorch_neuropod lets you specify python packages required by the model. If the dependency is not already available, it will be downloaded and unpacked (not quite installed; see below).

You also mentioned in code that this is problematic when running multiple python models in a single process and it's only intended to work when using OPE.

If you're running multiple python models without OPE, things can get really tricky (even if you're not using requirements). This is because, without OPE (out-of-process execution), all models will run in the same process using the same python interpreter.

It's possible that this can cause conflicts on its own (e.g. one model changing global state that another depends on), but it gets much more complex when requirements are introduced. If one model depends on torch==1.8.0 and one depends on torch==1.9.0, both can't be loaded into the same python interpreter (in a robust way).

Additionally, conflicts between transitive native dependencies of models (e.g. different versions of Eigen used by torch and tensorflow) can cause really really complicated problems.

Running with OPE solves these issues because every model runs in its own isolated environment independent of your application and independent of every other model.

So, I am wondering is it possible to pre-install the necessary package like torch to the isolated python environment before loading the model and keep the size of the python backend small at the same time?

Dependencies are only downloaded and unpacked the first time they're used and then are immediately available on subsequent model requests for the same dependency. One thing to note is that the packages aren't actually installed; the unpacked packages are added to sys.path when requested by a model. This lets us support multiple versions of packages (e.g torch 1.8 and 1.9) in a clean way. See here for more details.

If you control the runtime environment, you can preload dependencies during your environment build process by running a placeholder model that depends on everything you want to preload:

create_python_neuropod(
    ...
    requirements="""
    torch==1.8.0
    """
    ...
)

And as long as you persist the cache folder (~/.neuropod/pythonpackages/py{major}{minor}/), subsequent uses of torch==1.8.0 in that environment will not go through a download/unpacking process.

Final notes

If you can ensure that all the models you run won't have conflicting dependencies (or transitive dependencies), you might be able to get away with not using OPE, but that doesn't seem particularly robust.

Also, if you want specific packages to be available to all of your models (without them specifying requirements), you can add them here:

This triggers a download/unpack at runtime on the first run of any python model, but the difference here is that those packages are available to all models regardless of if they're specified in a model's requirements or not.

Let me know if you have any other questions!

from neuropod.

Hi Vivek, thanks for your response. This answered all my questions. But when I checked the requirements.lock file in my model, I still could see --index-url and --trusted-host in it which would cause ValueError when loading deps. I remembered that I fixed them in commit here. How can I the neuropod package that has my fix?

from neuropod.