Skip to content

Conversation

victoryc
Copy link
Contributor

Added support for GroupNorm.

Venkat Ajjanagadde and others added 9 commits April 2, 2021 18:11
…parameter happens to be on the gpu. The failure is caused due to the two parts being concatenated not being on the same device. By default, torch.ones creates the tensor on the cpu. So, if the coord happens to be on gpu, the error occurs. Fixed by creating the torch.ones tensor on the same device as the coord.
…e chances of bucket overflow error happening in hashmap.cu. When n is small, the effect of uneven distribution of hash values is more pronounced, and we select the table size to be 4 * nextPow2 and for larger n, we take it to be 2 * nextPow2. This way, for larger n, when memory size consideration is more important, the table_size gets computed to be same as it used to be.
…parameter happens to be on the gpu. The failure is caused due to the two parts being concatenated not being on the same device. By default, torch.ones creates the tensor on the cpu. So, if the coord happens to be on gpu, the error occurs. Fixed by creating the torch.ones tensor on the same device as the coord.
…e chances of bucket overflow error happening in hashmap.cu. When n is small, the effect of uneven distribution of hash values is more pronounced, and we select the table size to be 4 * nextPow2 and for larger n, we take it to be 2 * nextPow2. This way, for larger n, when memory size consideration is more important, the table_size gets computed to be same as it used to be.
@zhijian-liu zhijian-liu self-requested a review April 29, 2021 19:57
Copy link
Contributor

@zhijian-liu zhijian-liu left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the PR!

@zhijian-liu zhijian-liu merged commit e07e12f into mit-han-lab:master Apr 29, 2021
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants