Replies: 1 comment 2 replies
-
|
Yes, this will probably the next major refactoring. I was planning to have already started working on this, but got sidetracked with a few Metal optimizations the last couple of days. It's next on the list now. |
Beta Was this translation helpful? Give feedback.
-
|
Great to hear! Do you mind if I propose a PR to that end or do you want to do it entirely yourself? |
Beta Was this translation helpful? Give feedback.
-
|
I'll try to start a draft PR in the next days - we already have had some discussions how to approach the refactoring so it would be easier if I initiate it. We can discuss when it is open. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Maybe it's a stupid question or maybe it's already been answered before, but: is there any particular reason why
llama-model.cppis kept as such a monolithic beast instead of moving all the graph builders to their separate files in somemodels/subdirectory?Obviously such a huge file is rough for IDE parsers and stuff, but the place where it really gets bad is using LLM assistants. I sometimes try to use i.e. Gemini to compare new model implementations with their reference Python implementation, but even Gemini with its 1 mil context struggles with analyzing the entire
llama-model.cpp.Would it be possible to actually refactor the file and move all the builders out or is there a good reason for why it stays as it is?
Beta Was this translation helpful? Give feedback.
All reactions