diff --git a/benchmarks/README.md b/benchmarks/README.md
index 716931c..3e82680 100644
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -28,8 +28,9 @@ Provide run.py Python script with following arguments:
 - -p, prompt size(s) to benchmark, size of an input prompt; multiple prompt sizes can be provided and they will be treated as separate cases to benchmark
 - -r, thread-range, e.g., on an 80-thread system, it should be input as 0-79, unless user wants to use just a subset of available threads, say 16-63 (48 threads indexed 16<>63)
 - -fa, 0/1, disable/enable flash attention, default: 0
+- -d, docker_image, docker image used for benchmarking, default: amperecomputingai/llama.cpp:latest
 ```bash
-python3 run.py -m Meta-Llama-3-8B-Instruct.Q8_0.gguf -t 10 16 32 40 64 80 -b 1 2 4 8 16 32 64 -p 512 -r 0-79 -fa 1
+python3 run.py -m Meta-Llama-3-8B-Instruct.Q8_0.gguf -t 10 16 32 40 64 80 -b 1 2 4 8 16 32 64 -p 512 -r 0-79 -fa 1 -d amperecomputingai/llama.cpp:3.2.1-ampereone
 ```
 
 ## Quick run on 80t OCI A1 system
diff --git a/benchmarks/run.py b/benchmarks/run.py
index fd58448..a09978b 100644
--- a/benchmarks/run.py
+++ b/benchmarks/run.py
@@ -6,24 +6,24 @@
 import subprocess
 from utils.benchmark import parse_threads_range
 
-
 def get_file_dir():
     return os.path.dirname(os.path.realpath(__file__))
 
-
 def docker_init(node):
-    tag = "amperecomputingai/llama.cpp:3.1.2"
-    if subprocess.run(
-            ["docker", "pull", tag]).returncode != 0:
-        print("Docker pull process failed!")
-        sys.exit(1)
+
+    if subprocess.run(["docker", "inspect", docker_image], capture_output=True).returncode != 0:
+        print(f"Docker image {docker_image} doesn't exsit, try to pull it.")
+        if subprocess.run(["docker", "pull", docker_image], capture_output=True).returncode != 0:
+            print("Docker pull process failed!")
+            sys.exit(1)
+
     container_name = f"llama_benchmark_n{node}"
-    subprocess.run(["docker", "rm", "-f", container_name])
+    subprocess.run(["docker", "rm", "-f", container_name], capture_output=True)
     memory = (psutil.virtual_memory().total >> 30) - 30  # leave 30GB for OS
     assert memory > 10, "less than 10GB of memory available on the system for llama.cpp"
     if subprocess.run(
             ["docker", "run", "--privileged=true", "--cpuset-mems", f"{str(node)}", "--name", container_name, "-d", "-m", f"{str(memory)}g", "-v",
-             f"{get_file_dir()}:/runner", "--entrypoint", "/bin/bash", "-it", tag]).returncode != 0:
+             f"{get_file_dir()}:/runner", "--entrypoint", "/bin/bash", "-it", docker_image],capture_output=True).returncode != 0:
         print("Docker run process failed!")
         sys.exit(1)
     return container_name
@@ -33,13 +33,13 @@ def docker_restart(docker_name):
     break_time = 15
 
     def docker_stop():
-        if subprocess.run(["docker", "stop", docker_name]).returncode != 0:
+        if subprocess.run(["docker", "stop", docker_name], capture_output=True).returncode != 0:
             print(f"Stopping docker container {docker_name} failed, retrying in {break_time} seconds.")
             time.sleep(break_time)
             docker_stop()
 
     def docker_start():
-        if subprocess.run(["docker", "start", docker_name]).returncode != 0:
+        if subprocess.run(["docker", "start", docker_name], capture_output=True).returncode != 0:
             print(f"Starting docker container {docker_name} failed, retrying in {break_time} seconds.")
             time.sleep(break_time)
             docker_start()
@@ -94,6 +94,9 @@ def parse_args():
     parser.add_argument("-m", "--model_names",
                         type=str, required=True, nargs="+",
                         help="model names, e.g. 'Meta-Llama-3-8B-Instruct.Q8_0.gguf'")
+    parser.add_argument("-d", "--docker_image",
+                        type=str, required=True,
+                        help="Docker image to use for benchmarking")
     parser.add_argument("-t", "--num_threads",
                         type=int, required=True, nargs="+",
                         help="number of threads per process to use")
@@ -121,7 +124,11 @@ def parse_args():
 
 
 def main():
+    global docker_image
+
     args = parse_args()
+    docker_image = args.docker_image
+    print(f"Test with docker image {docker_image}")
     benchmark(docker_init(args.numa), args)