Skip to content

Conversation

@noorbhatia
Copy link
Contributor

Fix #47

@noorbhatia noorbhatia marked this pull request as draft November 27, 2025 18:00
@noorbhatia
Copy link
Contributor Author

Crashing with RESOURCE_TYPE_MEMORY: high watermark memory limit exceeded

@mattt
Copy link
Owner

mattt commented Dec 3, 2025

@noorbhatia Thanks for your work on this!

Crashing with RESOURCE_TYPE_MEMORY: high watermark memory limit exceeded

Interesting. Was this happening at all without this change? Looking at this MLX issue, it looks like the issue stems from unbounded growth of the KV cache, and can be addressed by either configuring a reasonable limit or resetting after processing the prompt.


Aside from that, is there anything more to do with the PR? Or is this ready for review?

@noorbhatia noorbhatia marked this pull request as ready for review December 9, 2025 15:35
@noorbhatia
Copy link
Contributor Author

Hey @mattt , apologies for the delay.

I was able to test it more thoroughly and I think it's ready for review.

@noorbhatia noorbhatia force-pushed the noor/fix-mlx-instructions branch from 43014bd to 1500acc Compare December 10, 2025 16:58
@mattt
Copy link
Owner

mattt commented Dec 11, 2025

Fantastic work, @noorbhatia! I just updated the README to improve instructions for testing MLX. Following those, everything seems to be working as expected. Merging this now.

@mattt mattt merged commit 877ee49 into mattt:main Dec 11, 2025
3 checks passed
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.

Support .system role for instructions in MLXLanguageModel

2 participants