diff --git a/src/transformers/testing_utils.py b/src/transformers/testing_utils.py
index a9ab304d2aa650..b93a40daa2ac1b 100644
--- a/src/transformers/testing_utils.py
+++ b/src/transformers/testing_utils.py
@@ -175,6 +175,7 @@ def parse_int_from_env(key, default=None):
 _tf_gpu_memory_limit = parse_int_from_env("TF_GPU_MEMORY_LIMIT", default=None)
 _run_pipeline_tests = parse_flag_from_env("RUN_PIPELINE_TESTS", default=True)
 _run_tool_tests = parse_flag_from_env("RUN_TOOL_TESTS", default=False)
+_run_third_party_device_tests = parse_flag_from_env("RUN_THIRD_PARTY_DEVICE_TESTS", default=False)
 
 
 def is_pt_tf_cross_test(test_case):
@@ -612,7 +613,12 @@ def require_torch_multi_npu(test_case):
     # Set env var CUDA_VISIBLE_DEVICES="" to force cpu-mode
     import torch
 
-    torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+    if torch.cuda.is_available():
+        torch_device = "cuda"
+    elif _run_third_party_device_tests and is_torch_npu_available():
+        torch_device = "npu"
+    else:
+        torch_device = "cpu"
 else:
     torch_device = None