NVIDIA / NVIDIA/Megatron-LM

moe-token-dispatcher-type alltoall error

Open
#1,877 2 comments 0 reactions 1 assignee Claimed by @yashaswikarnati View on GitHub
bug community-request module: moe waiting-on-maintainers
Dominant language
Python
Stars
17.9k
Forks
4.5k
Avg merge
4d 6h
Merged PRs (30d)
271

Description

**Describe the bug**

Got runtime error when running moe megatron training on Ampere with:
--tensor-model-parallel-size 4
--expert-model-parallel-size 8
--expert-tensor-parallel-size 1
--moe-token-dispatcher-type alltoall

**Steps/Code to reproduce bug**

`[rank11]: output = output.view(self.hidden_shape)
[rank11]: RuntimeError: shape '[2048, 1, 4096]' is invalid for input of size 0`

https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/transformer/moe/token_dispatcher.py#L803

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.