[projects][vllm] Reenable vLLM

.gitignore

@@ -6,6 +6,7 @@ work/
 airootfs/root/customize_airootfs.sh
 airootfs/root/customize_airootfs/scripts/1000-automatic.sh
 airootfs/root/customize_airootfs/scripts/1000-comfyui.sh
+airootfs/root/customize_airootfs/scripts/1000-invokeai.sh
 airootfs/root/customize_airootfs/scripts/1000-llamacpp.sh
 airootfs/root/customize_airootfs/scripts/1000-sillytavern.sh
 airootfs/root/customize_airootfs/scripts/1000-vllm.sh

airootfs/root/customize_airootfs/scripts/1000-vllm.sh.jinja2

@@ -0,0 +1,16 @@
+#!/bin/sh
+set -eu
+
+{% if platform == "cuda" %}
+# use patched vllm
+export PIP_EXTRA_INDEX_URL="https://sasha0552.github.io/pascal-pkgs-ci/"
+
+# install vllm
+pipx install vllm
+
+# remove triton
+pipx runpip vllm uninstall -y triton
+
+# install patched triton
+pipx runpip vllm install triton-pascal
+{% endif %}