Open mlondschien opened 6 days ago
In general, we support linux-arm64
(at least within the conda-forge ecosystem). For us, the most common use case is running a linux docker image on Apple silicon.
What page size is your system running on?
getconf PAGESIZE
# 4096
Just googling your problem, a similar issue popped up here: https://github.com/pola-rs/polars/issues/15549. Setting JEMALLOC_SYS_WITH_LG_PAGE=16
when building the package might help.
cc @xhochy
I opened a PR here: https://github.com/conda-forge/tabmat-feedstock/pull/50
Thanks Jan!
mlondsch@nid005009:~/code/$ getconf PAGESIZE
65536
A quick
export JEMALLOC_SYS_WITH_LG_PAGE=16
pip install tabmat-4.0.1.tar.gz --no-deps
raised an error, I'll have a closer look tomorrow.
@mlondschien can you try the latest build tabmat-4.0.1-py312ha2895bd_1
from conda-forge? Note that the build number was bumped to 1.
Same as before.
Following the installation steps of tabmat
in
https://github.com/Quantco/tabmat/blob/e288b47bfbf04c92ce824008f0cf615d32bcbad7/pyproject.toml#L81-L95
(as suggested by the awesome cscs support)
git clone --branch 5.3.0 https://github.com/jemalloc/jemalloc.git
cd jemalloc
./autogen.sh --disable-cxx --with-jemalloc-prefix=local --with-install-suffix=local --disable-initial-exec-tls
make
make install_bin install_include install_lib
ldconfig
cd ../
git clone --branch 12.1.1 https://github.com/xtensor-stack/xsimd.git
cd xsimd/
cmake -B build .
cmake --build build -t install
cd ../
git clone https://github.com/Quantco/tabmat.git
cd tabmat/
pip install --use-pep517 .
# Test that it works
python -c 'import glum; import numpy as np; X = np.random.rand(10, 3); y = np.random.rand(10);g = glum.GeneralizedLinearRegressor().fit(X, y); print(g.coef_table())'
intercept -0.234698
_col_0 -0.054386
_col_1 1.107723
_col_2 0.351426
Name: coef, dtype: float64
to build locally did the trick!
Did you set the JEMALLOC_SYS_WITH_LG_PAGE
environment variable before building this?
No. It's also not set by default.
mlondsch@nid005009:~/code$ echo $JEMALLOC_SYS_WITH_LG_PAGE
edit: below are the full build logs, in case this helps
I am using
glum
on anarm64
machine. Both after installing fromconda
andpip
(NB: why are there no arm64 linux wheels? Outputs below are fromconda
installation) and importingglum
, I getFitting a model results in a segmentation fault:
Does glum support arm64 chips?