diff --git a/.github/scripts/e2e-test-linux-and-mac.sh b/.github/scripts/e2e-test-linux-and-mac.sh
index 0d850ea48..1c5fdf7f4 100644
--- a/.github/scripts/e2e-test-linux-and-mac.sh
+++ b/.github/scripts/e2e-test-linux-and-mac.sh
@@ -14,8 +14,14 @@ rm /tmp/response1.log /tmp/response2.log /tmp/nitro.log
 BINARY_PATH=$1
 DOWNLOAD_URL=$2
 
+# Random port to ensure it's not used
+min=10000
+max=11000
+range=$((max - min + 1))
+PORT=$((RANDOM % range + min))
+
 # Start the binary file
-"$BINARY_PATH" 1 127.0.0.1 5000 > /tmp/nitro.log 2>&1 &
+"$BINARY_PATH" 1 127.0.0.1 $PORT > /tmp/nitro.log 2>&1 &
 
 # Get the process id of the binary file
 pid=$!
@@ -29,15 +35,13 @@ fi
 # Wait for a few seconds to let the server start
 sleep 5
 
-
-
 # Check if /tmp/testmodel exists, if not, download it
 if [[ ! -f "/tmp/testmodel" ]]; then
     wget $DOWNLOAD_URL -O /tmp/testmodel
 fi
 
 # Run the curl commands
-response1=$(curl -o /tmp/response1.log -s -w "%{http_code}" --location 'http://127.0.0.1:5000/inferences/llamacpp/loadModel' \
+response1=$(curl -o /tmp/response1.log -s -w "%{http_code}" --location "http://127.0.0.1:$PORT/inferences/llamacpp/loadModel" \
 --header 'Content-Type: application/json' \
 --data '{
     "llama_model_path": "/tmp/testmodel",
@@ -46,7 +50,7 @@ response1=$(curl -o /tmp/response1.log -s -w "%{http_code}" --location 'http://1
     "embedding": false
 }' 2>&1)
 
-response2=$(curl -o /tmp/response2.log -s -w "%{http_code}" --location 'http://127.0.0.1:5000/inferences/llamacpp/chat_completion' \
+response2=$(curl -o /tmp/response2.log -s -w "%{http_code}" --location "http://127.0.0.1:$PORT/inferences/llamacpp/chat_completion" \
 --header 'Content-Type: application/json' \
 --header 'Accept: text/event-stream' \
 --header 'Access-Control-Allow-Origin: *' \
diff --git a/.github/scripts/e2e-test-windows.bat b/.github/scripts/e2e-test-windows.bat
index 4389531c1..96a5385de 100644
--- a/.github/scripts/e2e-test-windows.bat
+++ b/.github/scripts/e2e-test-windows.bat
@@ -20,10 +20,15 @@ del %TEMP%\response1.log 2>nul
 del %TEMP%\response2.log 2>nul
 del %TEMP%\nitro.log 2>nul
 
+set /a min=9999
+set /a max=11000
+set /a range=max-min+1
+set /a PORT=%min% + %RANDOM% %% %range%
+
 rem Start the binary file
-start /B "" "%BINARY_PATH%" 1 "127.0.0.1" 5000 > %TEMP%\nitro.log 2>&1
+start /B "" "%BINARY_PATH%" 1 "127.0.0.1" %PORT% > %TEMP%\nitro.log 2>&1
 
-ping -n 6 127.0.0.1 5000 > nul
+ping -n 6 127.0.0.1 %PORT% > nul
 
 rem Capture the PID of the started process with "nitro" in its name
 for /f "tokens=2" %%a in ('tasklist /fi "imagename eq %BINARY_NAME%" /fo list ^| findstr /B "PID:"') do (
@@ -55,9 +60,9 @@ echo curl_data1=%curl_data1%
 echo curl_data2=%curl_data2%
 
 rem Run the curl commands and capture the status code
-curl.exe -o %TEMP%\response1.log -s -w "%%{http_code}" --location "http://127.0.0.1:5000/inferences/llamacpp/loadModel" --header "Content-Type: application/json" --data "%curl_data1%" > %TEMP%\response1_code.log 2>&1
+curl.exe -o %TEMP%\response1.log -s -w "%%{http_code}" --location "http://127.0.0.1:%PORT%/inferences/llamacpp/loadModel" --header "Content-Type: application/json" --data "%curl_data1%" > %TEMP%\response1_code.log 2>&1
 
-curl.exe -o %TEMP%\response2.log -s -w "%%{http_code}" --location "http://127.0.0.1:5000/inferences/llamacpp/chat_completion" ^
+curl.exe -o %TEMP%\response2.log -s -w "%%{http_code}" --location "http://127.0.0.1:%PORT%/inferences/llamacpp/chat_completion" ^
 --header "Content-Type: application/json" ^
 --header "Accept: text/event-stream" ^
 --header "Access-Control-Allow-Origin: *" ^
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 9cd93ab96..2e9a5b98d 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -41,7 +41,7 @@ jobs:
           prerelease: false
 
   ubuntu-amd64-build:
-    runs-on: linux-gpu
+    runs-on: ubuntu-latest
     needs: create-draft-release
     if: always() && (needs.create-draft-release.result == 'success' || needs.create-draft-release.result == 'skipped')
     permissions:
diff --git a/llama.cpp b/llama.cpp
index 1f5cd8327..1d144112c 160000
--- a/llama.cpp
+++ b/llama.cpp
@@ -1 +1 @@
-Subproject commit 1f5cd83275fabb43f2ae92c30033b384a3eb37b4
+Subproject commit 1d144112c0fbbb4ecc07dbcf4f05a380148bd6de