ColossalAI/colossalai/zero/low_level
LuGY d86ddd9b29
[hotfix] fix unsafe async comm in zero (#4404)
* improve stablility of zero

* fix wrong index

* add record stream
2023-08-11 15:09:24 +08:00
..
bookkeeping [hotfix] fix unsafe async comm in zero (#4404) 2023-08-11 15:09:24 +08:00
__init__.py [zero] reorganize zero/gemini folder structure (#3424) 2023-04-04 13:48:16 +08:00
_utils.py [zero] allow passing process group to zero12 (#4153) 2023-07-31 22:13:29 +08:00
low_level_optim.py [hotfix] fix unsafe async comm in zero (#4404) 2023-08-11 15:09:24 +08:00
readme.md [zero] support shard optimizer state dict of zero (#4194) 2023-07-31 22:13:29 +08:00

readme.md

Low Level ZeRO

Low Level ZeRO == ZeRO-DP stage 1 and 2, we would denote it as ZeRO.

Design:

Notion

p32 denotes the param copy in the optimizer p denotes the model param g denotes the gradient

INIT

In low level zero(1, 2), p32 is split. Different from the previous implement, we split each p32 evenly by world_size. Thus, rank0 got a param list as [p00, p10], rank1 got a param list as [p-01, p-11], etc. image

For the detailed implementation, we first pad p for it can be split by world_size if needed. Then, we would view it to the shape [world_size, -1], and each rank got its own part p32 by cloning.

BWD

To leverage the communication, a gradient would be added to a bucket first. When the bucket is full, each g in it would be reshaped as [world_size, -1]. And the [local_rank] parts would be united. The data structure looks like this:

{
0: [g-00, g-10],
1: [g-01, g-11],
2: [g-02, g-12]
}

After that, the gradients would be flattened by rank, and the data structure looks like this:

# g-0 means flatten([g-00, g-10])
{
0: [g-0],
1: [g-1],
2: [g-2]
}

For zero1, we iterate the dictionary and do all_reduce. For zero2, we can just do reduce-scatter.

Optim

For each rank gets its own p32 and the counterpart g, it is quite easy to do optim.step().

However, we have to consider a situation of layer drop, for instance:

class MlpModel(nn.Module):
    def __init__(self):
        super(MlpModel, self).__init__()
        self.linear1 = nn.Linear(128, 256)
        self.drop_linear = nn.Linear(256, 256)
        self.linear2 = nn.Linear(256, 512)

    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

And the solution is to build a mapping of p32, p, and g. Before optim.step(), we collect p which requires_grad=True and p.grad != None as a real working param. And select the counterpart p32 and g.