pytorch-dev.yaml

name: pytorch-dev
channels:
  - conda-forge
  - nodefaults
dependencies:
  - python=3.8
  - pip

  # Select compiler-version (optional, but recommended)
  # These are pulled in by cuda-nvcc implicitly
  - gcc=12
  - gxx=12

  # CUDA requirements, even if using system CUDA
  - magma
  - cuda-driver-dev
  - cuda-version=12.2
  - cudnn
  - conda-gcc-specs # Needed for triton to find libcuda.so

  # The following 6 packages install cudatoolkit
  # Not necessary in qgpu as it's installed system-wide
  - cuda-libraries-dev
  - cuda-nvcc
  - cuda-gdb
  - cuda-nvtx-dev
  - cuda-nvml-dev
  - cuda-cupti-dev

  # PyTorch dependencies
  - ccache
  - cmake
  - fsspec
  - mkl
  - mkl-include
  - ninja
  - packaging
  - pyyaml
  - scikit-build
  - setuptools
  - sysroot_linux-64>=2.17
  - types-dataclasses
  - typing
  - typing-extensions
  - wheel

  # Test dependencies
  - numpy
  - expecttest
  - hypothesis
  - lintrunner
  - scipy
  - psutil
  - pytest

  # Development utilities
  - ghstack
  - ipython
  - pytest-sugar
  - pytest-xdist

  # torch.compile, hard dependencies
  - filelock
  - sympy

  # torch.compile, optional dependencies
  - astunparse
  - dill
  - jinja2
  - lark
  - optree
  - tabulate
  - networkx
  - pip:
    - z3-solver

  # triton dependencies, if building from source
  - zlib

  # torchbench dependencies that may cause issues if installed by pip
  # - pandas
  # - transformers