MPMD detected error when using optimum-neuron with TP

[michaelbenayoun]

So basically I am trying to train LLama / Mistral.

I run the following command:

NEURON_RT_LOG_LEVEL=info XLA_USE_BF16=1 ./train_mistral.sh

Here is the link to train_mistral.sh

The issue is that I get MPMD detected. It means that at some point at least 2 workers try to execute different graphs. So I tried to check the diff between the two HLO graphs. I ran the script multiple times, I cannot say I always end-up with the same diff, but at least multiple times I ended up with this:

Basically:

• In one case, one parameter is the input to a select op.

• In the other case, we have 2 parameters: the same one as in the first case, let's call it p, and a scalar., let's call it s . Then the input to the select op is in this case: p - broadcast(s).

After analyzing it a bit, I think it this computation comes from the ParallelEmbedding layer. For some reason what is considered a constant equal to 0 in one case, is considered a parameter in the other case.

I thought it could be linked to scalar specialization by XLA so I also ran the job with XLA_NO_SPECIAL_SCALARS=1 but ended up with a MPMD detected error as well.

So I tried not to use ParallelEmbedding. When sequence parallelism is enabled I end-up with:

• In one case it does [16, 64] -> [1, 16, 64]. So here it seems to be B x S x H. Then it adds a reshape at the end to become S / 2 x B x H.
• And in the other case [1, 64] -> [16, 1, 64] . Here it is S x B x H. And then we end-up with S / 2 x B x H.

Finally, I tried disabling sequence parallelism and ended-up with:

Note: when I disable tensor parallelism it seems to be working properly.

[michaelbenayoun]

It was linked to torch.autocast.