Allow exclusion of torch GPU dependencies during installation

[spreeni]

Is your feature request related to a problem? Please describe.

When installing farm-haystack on Linux, torch is installed with GPU support. This causes the Haystack dependencies to go up from 1.4GB to 4GB which is limiting for deployment. On Mac and Windows this is not the case due to how torch dependencies are defined, for example in my poetry.lock:

[[package]]
name = "torch"
version = "1.13.1"
description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
category = "dev"
optional = false
python-versions = ">=3.7.0"
files = [
    {file = "torch-1.13.1-cp310-cp310-manylinux1_x86_64.whl", hash = "sha256:fd12043868a34a8da7d490bf6db66991108b00ffbeecb034228bfcbbd4197143"},
    {file = "torch-1.13.1-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:d9fe785d375f2e26a5d5eba5de91f89e6a3be5d11efb497e76705fdf93fa3c2e"},
    {file = "torch-1.13.1-cp310-cp310-win_amd64.whl", hash = "sha256:98124598cdff4c287dbf50f53fb455f0c1e3a88022b39648102957f3445e9b76"},
    {file = "torch-1.13.1-cp310-none-macosx_10_9_x86_64.whl", hash = "sha256:393a6273c832e047581063fb74335ff50b4c566217019cc6ace318cd79eb0566"},
    {file = "torch-1.13.1-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:0122806b111b949d21fa1a5f9764d1fd2fcc4a47cb7f8ff914204fd4fc752ed5"},
    {file = "torch-1.13.1-cp311-cp311-manylinux1_x86_64.whl", hash = "sha256:22128502fd8f5b25ac1cd849ecb64a418382ae81dd4ce2b5cebaa09ab15b0d9b"},
    {file = "torch-1.13.1-cp37-cp37m-manylinux1_x86_64.whl", hash = "sha256:76024be052b659ac1304ab8475ab03ea0a12124c3e7626282c9c86798ac7bc11"},
    {file = "torch-1.13.1-cp37-cp37m-manylinux2014_aarch64.whl", hash = "sha256:ea8dda84d796094eb8709df0fcd6b56dc20b58fdd6bc4e8d7109930dafc8e419"},
    {file = "torch-1.13.1-cp37-cp37m-win_amd64.whl", hash = "sha256:2ee7b81e9c457252bddd7d3da66fb1f619a5d12c24d7074de91c4ddafb832c93"},
    {file = "torch-1.13.1-cp37-none-macosx_10_9_x86_64.whl", hash = "sha256:0d9b8061048cfb78e675b9d2ea8503bfe30db43d583599ae8626b1263a0c1380"},
    {file = "torch-1.13.1-cp37-none-macosx_11_0_arm64.whl", hash = "sha256:f402ca80b66e9fbd661ed4287d7553f7f3899d9ab54bf5c67faada1555abde28"},
    {file = "torch-1.13.1-cp38-cp38-manylinux1_x86_64.whl", hash = "sha256:727dbf00e2cf858052364c0e2a496684b9cb5aa01dc8a8bc8bbb7c54502bdcdd"},
    {file = "torch-1.13.1-cp38-cp38-manylinux2014_aarch64.whl", hash = "sha256:df8434b0695e9ceb8cc70650afc1310d8ba949e6db2a0525ddd9c3b2b181e5fe"},
    {file = "torch-1.13.1-cp38-cp38-win_amd64.whl", hash = "sha256:5e1e722a41f52a3f26f0c4fcec227e02c6c42f7c094f32e49d4beef7d1e213ea"},
    {file = "torch-1.13.1-cp38-none-macosx_10_9_x86_64.whl", hash = "sha256:33e67eea526e0bbb9151263e65417a9ef2d8fa53cbe628e87310060c9dcfa312"},
    {file = "torch-1.13.1-cp38-none-macosx_11_0_arm64.whl", hash = "sha256:eeeb204d30fd40af6a2d80879b46a7efbe3cf43cdbeb8838dd4f3d126cc90b2b"},
    {file = "torch-1.13.1-cp39-cp39-manylinux1_x86_64.whl", hash = "sha256:50ff5e76d70074f6653d191fe4f6a42fdbe0cf942fbe2a3af0b75eaa414ac038"},
    {file = "torch-1.13.1-cp39-cp39-manylinux2014_aarch64.whl", hash = "sha256:2c3581a3fd81eb1f0f22997cddffea569fea53bafa372b2c0471db373b26aafc"},
    {file = "torch-1.13.1-cp39-cp39-win_amd64.whl", hash = "sha256:0aa46f0ac95050c604bcf9ef71da9f1172e5037fdf2ebe051962d47b123848e7"},
    {file = "torch-1.13.1-cp39-none-macosx_10_9_x86_64.whl", hash = "sha256:6930791efa8757cb6974af73d4996b6b50c592882a324b8fb0589c6a9ba2ddaf"},
    {file = "torch-1.13.1-cp39-none-macosx_11_0_arm64.whl", hash = "sha256:e0df902a7c7dd6c795698532ee5970ce898672625635d885eade9976e5a04949"},
]

[package.dependencies]
nvidia-cublas-cu11 = {version = "11.10.3.66", markers = "platform_system == \"Linux\""}
nvidia-cuda-nvrtc-cu11 = {version = "11.7.99", markers = "platform_system == \"Linux\""}
nvidia-cuda-runtime-cu11 = {version = "11.7.99", markers = "platform_system == \"Linux\""}
nvidia-cudnn-cu11 = {version = "8.5.0.96", markers = "platform_system == \"Linux\""}
typing-extensions = "*"

[package.extras]
opt-einsum = ["opt-einsum (>=3.3)"]

Describe the solution you'd like

Similar to how other extras during installation are defined like all/all-gpu, faiss/faiss-gpu, docstores/docstores-gpu, the default install pip install farm-haystack could only provide CPU-support, and pip install farm-haystack[gpu] would enable GPU-support.

Describe alternatives you've considered

We ran into this issue when developing on Mac, but deploying in a Linux docker container. If only CPU-support is wanted, an option working with poetry is adding the following CPU-only dependencies for torch and torchvision to your pyproject.toml.

torch = {url = "https://download.pytorch.org/whl/cpu/torch-1.13.1%2Bcpu-cp39-cp39-linux_x86_64.whl"}
torchvision = {url = "https://download.pytorch.org/whl/cpu/torchvision-0.14.1%2Bcpu-cp39-cp39-linux_x86_64.whl"}

This however makes the dependencies platform-specific, which is not ideal.

Additional context

This seems to be a general problem of torch with dependency management, as it does not offer GPU-support as e.g. an extra dependency. This has been discussed for example in this issue in the torch repository:

pytorch/pytorch#26340

[preritdas]

This is an important point. PyTorch GPU dependencies including NVIDIA cublas, cudnn etc. are very large, take time, and are unnecessary if running without a GPU. Our dependency is farm-haystack[faiss,ocr]==1.13.2 which triggers all kinds of GPU dependency downloads that are never used or taken advantage of given we're deploying on Cloud Run, which is entirely CPU-bound.