environment.yml is unsatisfiable as written (mixed channels + pinned builds) and conflicts with pip torch==1.13.1 vs conda PyTorch 2.0/cu118

[AllanK24]

Hi! I’m trying to create the environment from environment.yml and keep hitting solver failures. The file looks like an exported env with many exact build pins and it mixes channel stacks; additionally the pip: section pins torch==1.13.1 while the conda deps specify torchaudio==2.0.2, torchvision==0.15.2, and pytorch-cuda==11.8 — which implies PyTorch 2.0.x, not 1.13.

What I ran

# from repo root
conda env create -n qsam -f environment.yml           # also tried --solver classic

What happened

Conda fails to solve with a large set of packages “excluded by strict repo priority”, e.g.

LibMambaUnsatisfiableError: Encountered problems while solving:

• package _openmp_mutex-5.1-1_gnu is excluded by strict repo priority
• package libgcc-ng-11.2.0-h1234567_1 is excluded by strict repo priority
• package pillow-9.4.0-py311h6a678d5_0 is excluded by strict repo priority
• package python-3.11.4-h955ad1f_0 is excluded by strict repo priority
  ... (many similar lines)

Why I think this happens

Channel order vs pinned builds
channels: lists pytorch, nvidia, conda-forge, defaults, but many deps are pinned to defaults builds (e.g. py311h6a678d5_0). With strict channel priority, those defaults builds get excluded because conda-forge sits above defaults. The solver can’t mix them.

Inconsistent PyTorch stack

Conda deps: torchaudio=2.0.2, torchvision=0.15.2, pytorch-cuda=11.8 → this corresponds to PyTorch 2.0.x.
Pip deps: torch==1.13.1 and several nvidia-cu11 wheels → these conflict with the conda CUDA 11.8 stack.

Suggested fixes

Any of these would make the env reproducible:

Provide a lean, channel-coherent environment.yml (GPU), e.g.

name: finetune-sam
channels: [pytorch, nvidia, defaults]
dependencies:
  - python=3.11
  - pip
  - pytorch=2.0.1
  - torchvision=0.15.2
  - torchaudio=2.0.2
  - pytorch-cuda=11.8
  - numpy
  - scipy
  - scikit-image
  - pillow
  - matplotlib
  - jupyterlab
  - ffmpeg
  - pip:
      - segment-anything==1.0
      - monai==1.2.0
      - torchio==0.19.1
      - pytorch-lightning==2.0.9.post0
      - timm==0.9.7
      - transformers==4.26.1
      - tokenizers==0.13.3
      - simpleitk==2.2.1
      - opencv-python==4.8.0.76
      - pandas==2.1.0
      - tensorboard==2.14.0
      - safetensors==0.3.3
      - einops==0.6.1

(Key points: keep torch out of pip, let conda provide the CUDA build, and avoid exact build strings.)

• Or split into two files:
  environment.yml (conda-only, no pip torch); requirements-pip.txt (pip-only extras). In environment.yml, put defaults ahead of conda-forge or remove exact build pins so conda-forge can be first without strict-priority conflicts.

• If you want to keep conda-forge first: remove exact build pins (export with conda env export --from-history) so the solver can pick compatible conda-forge builds.

• Optional CPU-only variant (no NVIDIA dependency): drop nvidia channel and pytorch-cuda, and add cpuonly from pytorch if desired:

channels: [pytorch, defaults]
dependencies:
  - python=3.11
  - pip
  - pytorch=2.0.1
  - torchvision=0.15.2
  - torchaudio=2.0.2
  - cpuonly
  - ...

Environment

• Platform: linux-64
• Conda: (fill in) conda --version
• Channel priority: strict
• GPU/Driver: Driver Version: 550.163.01; CUDA Version: 12.4

I’m happy to open a PR with a cleaned environment.yml if this direction sounds good. Thanks!