New PRs not sychronized to main

SameDiff
1

seems the new 41 PRs have finished, why not merge them to main?

2

@Booker sorry just finished my last sprint here. I was finishing a massive number of memory leaks from the refactoring I was doing. I’m going through and doing extensive benchmark/testing for the internal rewrites. I wouldn’t say it’s done yet but I’ve done some benchmarking to make it a more solid foundation going forward.

1 Like
3

  • SameDiff Engine: Dynamic computation graphs

    • Define-by-run execution

    • Automatic gradient computation

    • Control flow support

    • Dynamic shapes // support sparse dimensions?

4

Expecting next release

5

@Booker yes of course. I’m still in the middle of adding a bunch of new features yet to modernize the framework yet. You can see what’s going on on my dev branch if you want: Dev branch to be segment by agibsonccc · Pull Request #10229 · deeplearning4j/deeplearning4j · GitHub

You’ll see

LORA

LLM finetuning

Onnx improvements

GGML loading

Cuda updates.

I’ll do what I can for the community but I’m doing what pays the bills for me first and foremost. This is what people want anyways.