Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

ggml : better way to express implicit node dependencies in a graph #502

Open
ggerganov opened this issue Sep 4, 2023 · 0 comments
Open
Labels
enhancement New feature or request refactoring Refactoring

Comments

@ggerganov
Copy link
Owner

Operations on views can introduce implicit dependencies between the nodes in the compute graph which often lead to bugs when we forget about these dependencies.

See ggerganov/llama.cpp#3012 for more context

We need to find some better way to handle such cases.
One option is to introduce ggml_depends_on() which if not ideal, should at least make the code more explicit less error prone.

@ggerganov ggerganov added enhancement New feature or request refactoring Refactoring labels Sep 4, 2023
CCLDArjun pushed a commit to CCLDArjun/ggml that referenced this issue Dec 18, 2023
* Retire the ggml_mul_mat() for transposed src0

- It can always be made contiguous with ggml_cpy()
- The code is now simplified
- The results are deterministic in respect to num threads

* SIMD-ify dequantize_row_q4_0() for ARM_NEON (ggerganov#502)

* Attempt to SIMD-ify dequantize_row_q4_0() for ARM_NEON

* Fix dequantization - forgot to interleave the quants
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
enhancement New feature or request refactoring Refactoring
Projects
Status: Todo
Development

No branches or pull requests

1 participant