Torch & double resolves, We have been struggling for days with getting torch projects build locally on Mac and on Linux Docker. I have read other posts about Torch, and the bugs, but I think our situation is slightly different as we have a bunch of existing Poetry projects that specify numeric requirements that fail with specific torch versions (+cpu, +cu118 etc). I think I have finally found something that works for our situation but I wonder if there is a way to make it more elegant? Here is what I have:

# Contains torch pinned to 2.0.0 but no cuda/nvidia dependencies

poetry_requirements(

name="reqs",

source="pyproject.toml",

module_mapping={

"imath": ["Imath"],

...

},

resolve="python-default"

)

# Contains all coda dependencies spelled out with version from a poetry.lock file

python_requirements(

name="cuda_reqs",

source="cuda_requirements.txt",

resolve="cuda",

)

# Getting dependencies again from the same Poetry file but overriding

# the Torch dependencies, as these are not picked up

# Uses another resolve since if I do it for python--local, I can't

# export a venv anymore because the coda dependencies are not found for Mac

poetry_requirements(

name="reqs_2",

source="pyproject.toml",

module_mapping={

"imath": ["Imath"],

...

},

overrides={

"torch": {

"dependencies": [

":cuda_reqs#nvidia-cublas-cu11",

...

":cuda_reqs#triton"

]

}

},

resolve="cuda"

)