[rpms/python-torch] rawhide: Improve paralellism on multisocket CPU

[rpms/python-torch] rawhide: Improve paralellism on multisocket CPU

[@ 2026-08-18 22:43 UTC (permalink / raw)]

            A new commit has been pushed.

            Repo   : rpms/python-torch
            Branch : rawhide
            Commit : 749d91ee2c96fc2d7e413d0cde294037e75f3f9a
            Author : Miroslav Suchý <msuchy@redhat.com>
            Date   : 2026-08-18T22:42:50+00:00
            Stats  : +1/-1 in 1 file(s)
            URL    : https://src.fedoraproject.org/rpms/python-torch/c/749d91ee2c96fc2d7e413d0cde294037e75f3f9a?branch=rawhide

            Log:
            Improve paralellism on multisocket CPU

Current formula counts CPUs in first socket only. This change counts
all physical cores in all sockets.

I have system with two socket (each of them has 28 cores. As result the job is
running with `-j 28` instead of `-j 112` and building at 25% utilization.

---
diff --git a/python-torch.spec b/python-torch.spec
index e9e522d..a0530ef 100644
--- a/python-torch.spec
+++ b/python-torch.spec
@@ -632,7 +632,7 @@ sed -i -e 's@USE_MSLK_DEFAULT ON@USE_MSLK_DEFAULT OFF@' CMakeLists.txt
 #
 %ifarch x86_64
 # Real cores, No hyperthreading
-COMPILE_JOBS=`cat /proc/cpuinfo | grep -m 1 'cpu cores' | awk '{ print $4 }'`
+COMPILE_JOBS=`lscpu -p=CORE | grep -v '^#' | sort -u |wc -l`
 %else
 # cpuinfo format varies on other arches, fall back to nproc
 COMPILE_JOBS=`nproc`