`accelerate` in `bloom-inference-scripts`?

[jeromeku]

Read the original blogpost comparing accelerate with deepspeed and am currently trying to benchmark using the bloom-inference-scripts. I'm a bit confused, as the accelerate script does not use accelerate library nor the accelerate launcher?

[jeromeku]

Nevermind :) Traced through the model loading code and see that setting device_map to "auto" results in memory-efficient model loading as implemented in accelerate.