fix checking GPU on ARM

images/benchmark-llm/benchmark.py

@@ -101,6 +101,11 @@
 def get_llama_cpp_path():
     """Check if GPU/CUDA is available, if not, use CPU-build of llama.cpp."""
     llama_cpp_path = "/llama_cpp_gpu"
+    # ARM64 doesn't have the CUDA binaries, so we need to use the CPU-build with early return
+    if not path.exists(llama_cpp_path):
+        logger.info("Using CPU-build of llama.cpp")
+        return "/llama_cpp_cpu"
+    # check if we can start the CUDA binary
     result = run(["./llama-cli", "--version"], cwd=llama_cpp_path, capture_output=True)
     if result.returncode != 0:
         llama_cpp_path = "/llama_cpp_cpu"