outlines
outlines copied to clipboard
dottxt-ai
Numpy > 1.26.4
What behavior of the library made you think about the improvement?
This issue is just meant as a Q&A, as I couldnt find anything specifically on this.
The question is why there is a dependency on older versions of numpy? What is the limitation in the newer numpy versions? And how big of a job do you estimate it would be to support newer versions of numpy?
How would you like it to behave?
Numpy > 1.26.4
@Mikeriess I don't think it would be a substantial effort.
We pinned to numpy 1 due to
- https://github.com/outlines-dev/outlines/issues/976
- https://github.com/outlines-dev/outlines/pull/977
The imports would need to be fixed. Unless there's something I'm missing this is a trivial change.
Got it. Have had some compatibility issues with outlines and the newest llama.cpp version due to this, so if this is a trivial change that is great news 👍
Got it. Have had some compatibility issues with outlines and the newest llama.cpp version due to this, so if this is a trivial change that is great news 👍
Is the latest llama.cpp incompatible with numpy<2.0.0? If so, could you please share your error?
I'd like to ensure this is high priority if it breaks the latest llama.cpp.
https://github.com/abetlen/llama-cpp-python/blob/9769e5719ad45d4be6bc3ebe01d0f568d3d5001e/pyproject.toml#L16
https://github.com/vllm-project/vllm/blob/5231f0898e559671c6c8cc48efc53a859fce1841/requirements-common.txt#L3
looks like everyone is lagging behind
Was the intention to allow importing numpy across all flavors of outlines? A few numpy<2 constraints remain in the optional packages, such as transformers: https://github.com/dottxt-ai/outlines/blob/main/pyproject.toml#L50-L54
On Windows pip wants to compile numpy 1.26.4 for some reason, and failing because I don't have a C++ dev environment. numpy 2.x got installed just fine.
