From 6482dd123dcd4b2d00052aa439b7da73f00048d1 Mon Sep 17 00:00:00 2001
From: Hyperling <me@hyperling.com>
Date: Wed, 21 Oct 2026 05:43:37 -0700
Subject: [PATCH] llm, change ollama_threads to use a better method of
 determining the number of cores to test, also fix output.

---
 tasks/workstation/linux/software/llm.yml | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/tasks/workstation/linux/software/llm.yml b/tasks/workstation/linux/software/llm.yml
index 4d65e20..1fd8aea 100644
--- a/tasks/workstation/linux/software/llm.yml
+++ b/tasks/workstation/linux/software/llm.yml
@@ -922,11 +922,16 @@
               context=',"num_ctx":'$num_ctx''
             fi
 
+            # Physical CPU Count (no longer using ansible_processor_vcpus, {{ ansible_processor_vcpus }})
+            # Will be something like 0-7, then get expanded to 0 1 2 3 4 5 6 7.
+            # Runnning with 0 causes default behavior, 1 per physical core, 8!
+            cores=$(cat /sys/devices/cpu_core/cpus | tr ',' '\n' | sed 's/-/ /' | xargs -L1 seq)
+
             echo " Stopping all ollama usage..."
             sudo sv restart ollama >/dev/null 2>&1
 
             # Benchmark Thread Count
-            for t in {1..{{ ansible_processor_vcpus | int }}}; do
+            for t in $cores; do
               start=$SECONDS
               echo -n "`date` - Model $model w/ $t Thread(s): "
 
@@ -938,7 +943,7 @@
               ctx="`ollama ps | grep $model | awk '{print $7}'`"
               time=$(( $finish - $start ))
 
-              echo ", tokens/sec: $speed @ context=$ctx, time: $time secs"
+              echo "tokens/sec: $speed @ context=$ctx, time: $time secs"
 
               ollama stop $model >/dev/null 2>&1
             done