[Discussion] About 3D Parallelism about colossalai HOT 6 CLOSED

Hi, @feifeibear . Thank you for sharing the idea! In our opinion, this is basically a trade-off between memory cost and communication cost.

The current design of 3D Linear layer applies an all-gather on the input matrix A and a reduce-scatter on the output matrix C in the forward pass (all-gather on the gradients of C and reduce-scatter on the gradients of A in the backward pass), so that each activation can be 1/N of the original size.

An alternative design is to use an all-reduce on C in the forward pass as well as on the gradients of A in the backward pass, but the activations are 1/N^(2/3) of the original size.

Considering activation checkpointing, as the forward pass is recomputed, the first design applies 2 * all-gather + reduce-scatter on A and 2 * reduce-scatter + all-gather on C in total, while the second design applies 3 * all-reduce. Since all-reduce of the ring algorithm has similar cost to all-gather + reduce-scatter, the total communication costs of both designs seem to be similar.

However, we indeed concern that small tensors decrease the bandwidth utilization, and it is hard to fuse them up. To find the optimal performance, we are testing as much models and networking environments as possible, and let the results tell.

from colossalai.

I agree 3D parallel can shrink the peak activation footprint in one GPU at cost of more communication. The method definitely works in some special cases. Maybe a simple searching method can be derived to figure out which part of the DNN is suitable for 3D parallelism in the constraint of a limited memory budget.

from colossalai.

I agree 3D parallel can shrink the peak activation footprint in one GPU at cost of more communication. The method definitely works in some special cases. Maybe a simple searching method can be derived to figure out which part of the DNN is suitable for 3D parallelism in the constraint of a limited memory budget.

This can be a good idea. For example, self-attention blocks usually consume more than mlp (ffn) blocks.

from colossalai.

This issue is stale because it has been open for 14 days with no activity.

from colossalai.

@1SAA communication profiling results may support some of my assumption iin discussion.

from colossalai.

We have updated a lot. This issue was closed due to inactivity. Thanks.

from colossalai.