Simplify torchserve installation procedure by removing the need to run the install_dependencies script

[agunapal]

🚀 The feature

Motivation

To install torchserve today, users have to run python ts_scripts/dependencies.py with the optional CUDA arg Ex: --cuda=cu102 and then install torchserve binaries on top. \

Example:

python ./ts_scripts/install_dependencies.py --cuda=cu102
pip install torchserve torch-model-archiver torch-workflow-archiver

The additional step of installing dependencies makes it confusing and its the responsibility of the user to run this additional step.

We want to automate this such that a single pip install or conda install command is sufficient to install torchserve ( including CUDA)

Current TorchServe Installation

PyPI

python ./ts_scripts/install_dependencies.py --cuda=cu102
pip install torchserve torch-model-archiver torch-workflow-archiver

Conda

python ./ts_scripts/install_dependencies.py --cuda=cu102
conda install -c pytorch torchserve torch-model-archiver torch-workflow-archiver

Proposed TorchServe Installation

PyPI

Specify dependencies like CUDA version, dev as extras to the pip install command

pip install torchserve[cu102] torch-model-archiver[cu102] torch-workflow-archiver[cu102]
pip install torchserve[cu102, dev] torch-model-archiver[cu102, dev] torch-workflow-archiver[cu102, dev]

Conda

conda install -c pytorch torchserve torch-model-archiver torch-workflow-archiver cudatoolkit=10.2

Test

PyPI

Conda

Alternatives

No response

Additional context

The following tasks need to be completed

• Common Tasks
  • Modify/cleanup the requirements.txt
• PyPI
  • Set up Github action with for the build process with a CUDA enabled machine and we need to switch between different versions of CUDA and build binaries.
  • Update Release scripts
  • Update Nightly scripts
  • Update README to support existing method for older binaries and new method for the newer ones
  • For the existing binaries/ older versions of torchserve, we would need to continue to support the existing way of installing
• Conda
  • Setup mutex package as described here
  • Set up Github action with for the build process with a CUDA enabled machine and we need to switch between different versions of CUDA and build binaries.
  • Update Release scripts
  • Update Nightly scripts
  • Update README to support existing method for older binaries and new method for the newer ones
  • For the existing binaries/ older versions of torchserve, we would need to continue to support the existing way of