Haven't the issues conda et al were trying to solve been solved with wheels and pip?
I've been supporting pytorch + cuda just using 'python3 -m venv foo' and pip (+ pip-compile) for the last year and it works - not super fast, but I prefer using the python.org blessed tooling.
Haven't the issues conda et al were trying to solve been solved with wheels and pip?
I've been supporting pytorch + cuda just using 'python3 -m venv foo' and pip (+ pip-compile) for the last year and it works - not super fast, but I prefer using the python.org blessed tooling.
Also, from a hot pip cache, on Ubuntu installing lighting takes ~45s and the venv is 5.2G. On an m1 mac, this takes ~17s and is 400MB.
Torch, nvidia, and triton site pkgs are quite a bit bigger on Ubuntu since I guess nvidia gpus aren't available on mac.