torchx icon indicating copy to clipboard operation
torchx copied to clipboard

dist.accelerate component

Open d4l3k opened this issue 11 months ago • 0 comments

This is a component that uses HuggingFace's accelerate in a similar manner to torchrun. I haven't tested with a accelerate config.

Fixes #834

Test plan:

pytest torchx/components/test/dist_test.py
scripts/component_integration_tests.py --scheduler local_cwd
torchx run -s local_cwd dist.accelerate -j2x2 --script torchx/examples/apps/compute_world_size/main.py

d4l3k avatar Mar 01 '24 00:03 d4l3k