diff --git a/.gitea/workflows/docker-build-push.yml b/.gitea/workflows/docker-build-push.yml
index dfa60e8..70c2fab 100644
--- a/.gitea/workflows/docker-build-push.yml
+++ b/.gitea/workflows/docker-build-push.yml
@@ -73,6 +73,8 @@ jobs:
             -t "$cambricon_image" .
           docker build \
             --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/enginex-metax/vllm:0.9.1 \
+            --build-arg PYTHON_BIN=/opt/conda/bin/python3 \
+            --build-arg VLLM_BIN=/opt/conda/bin/vllm \
             -t "$metax_image" .
           docker build \
             --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-vastai-vllm-vacc-fix-tokenizer:VVI-26.02 \
diff --git a/Dockerfile b/Dockerfile
index ab57b2c..14da321 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,12 +1,17 @@
 ARG BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0
 FROM ${BASE_IMAGE}
 
+ARG PYTHON_BIN=python3
+ARG VLLM_BIN=vllm
+ENV MODELHUB_COMPAT_PYTHON=${PYTHON_BIN}
+
 COPY engine_compat.py /opt/modelhub/engine_compat.py
 COPY vllm_wrapper.sh /opt/modelhub/vllm_wrapper.sh
 
 RUN set -eu; \
-    python3 -m py_compile /opt/modelhub/engine_compat.py; \
-    vllm_bin="$(command -v vllm)"; \
+    "$PYTHON_BIN" -m py_compile /opt/modelhub/engine_compat.py; \
+    if [ "$VLLM_BIN" = vllm ]; then vllm_bin="$(command -v vllm)"; else vllm_bin="$VLLM_BIN"; fi; \
+    test -x "$vllm_bin"; \
     mv "$vllm_bin" "$(dirname "$vllm_bin")/vllm_real"; \
     cp /opt/modelhub/vllm_wrapper.sh "$vllm_bin"; \
     chmod 0755 "$vllm_bin" /opt/modelhub/engine_compat.py
diff --git a/README.md b/README.md
index 7fa6521..8bd59bb 100644
--- a/README.md
+++ b/README.md
@@ -43,6 +43,8 @@ docker build --build-arg BASE_IMAGE=harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0
 docker build --build-arg BASE_IMAGE=harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260228 \
   -t qwen25-compat:mlu370 .
 docker build --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/enginex-metax/vllm:0.9.1 \
+  --build-arg PYTHON_BIN=/opt/conda/bin/python3 \
+  --build-arg VLLM_BIN=/opt/conda/bin/vllm \
   -t qwen25-compat:metax .
 docker build --build-arg BASE_IMAGE=harbor.4pd.io/modelhubxc/sunruoxi/enginex-vastai-vllm-vacc-fix-tokenizer:VVI-26.02 \
   -t qwen25-compat:vastai .
diff --git a/vllm_wrapper.sh b/vllm_wrapper.sh
index 86572f5..81208de 100644
--- a/vllm_wrapper.sh
+++ b/vllm_wrapper.sh
@@ -2,6 +2,7 @@
 set -euo pipefail
 
 vllm_real="$(dirname "$(readlink -f "$0")")/vllm_real"
+python_bin="${MODELHUB_COMPAT_PYTHON:-python3}"
 
 if [[ "${1:-}" != "serve" || -z "${2:-}" ]]; then
     exec "$vllm_real" "$@"
@@ -22,9 +23,9 @@ for ((i=0; i<${#args[@]}; i++)); do
 done
 
 if [[ -n "$tokenizer_dir" && -d "$tokenizer_dir" ]]; then
-    prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir" "$tokenizer_dir")"
+    prepared="$("$python_bin" /opt/modelhub/engine_compat.py "$model_dir" "$tokenizer_dir")"
 else
-    prepared="$(python3 /opt/modelhub/engine_compat.py "$model_dir")"
+    prepared="$("$python_bin" /opt/modelhub/engine_compat.py "$model_dir")"
     tokenizer_dir=""
 fi
 if [[ -z "$prepared" ]]; then