ColossalAI
ColossalAI copied to clipboard
[autoparallel] GPT2MLP runtime test
What does this PR do
- This PR construct a test to check the result of MLP with the auto parallel strategy equals to the original result or not.
- Update runtime preparation passes: a. add strategies_constructor as a optional argument to get the no strategy nodes list. b. Change the get_attr parameter shard workflow to fix a bug which makes parameters retain None as grad.
Note: In future PR, we need to fix the rng state issue for modules with randomness, such as nn.Dropout.