I was wondering how to perform multi-node inference? I'm very interested in running some experiments on Bloom (specifically probing its ability to reason). The high-level descriptions of Zero and DeepSpeed Inference indicate that multi-node inference is supported, but the examples I've found so far are only of multi-node training.
I was able to get deepspeed to perform inference instead of training in the multi-node setting on SLURM, but I ran into a different error. I'll open a different issue for that (#318).
I was wondering how to perform multi-node inference? I'm very interested in running some experiments on Bloom (specifically probing its ability to reason). The high-level descriptions of Zero and DeepSpeed Inference indicate that multi-node inference is supported, but the examples I've found so far are only of multi-node training.
I really appreciate any tips or pointers!