Skip to content

Conversation

@fshp971
Copy link

@fshp971 fshp971 commented Dec 28, 2023

What does this PR do?

Replace torch.Tensor with torch.empty when initializing model parameters.

The reason is that torch.empty is compatible with torch.set_default_device (but torch.Tensor is not), thus enable one to efficiently initialize a large ML model with specified GPUs.

(An example is that initialize a Llama2 model with the default torch.nn.init.xavier_normal_ method is extremely slow on CPUs, but very fast on GPUs.)

Before submitting

  • Did you have fun?
    • Make sure you had fun coding 🙃
  • Did you read the contributor guideline?
  • Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
    • N/A
  • Did you make sure to update the docs?
    • N/A
  • Did you write any new necessary tests?
    • N/A
  • Did you update the changelog? (if needed)
    • N/A

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

…llel/layers.py` to support `torch.set_default_device`
@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Dec 28, 2023
@mhjacobson
Copy link

I just hit this myself, and while I am not an expert in this particular project, I think it would be a good idea.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants