diff --git a/.github/workflows/self-push.yml b/.github/workflows/self-push.yml
index e21520d0e8..0bb5dbe87c 100644
--- a/.github/workflows/self-push.yml
+++ b/.github/workflows/self-push.yml
@@ -40,7 +40,8 @@ jobs:
 
     - name: Run all non-slow tests on GPU
       env:
-        TF_FORCE_GPU_ALLOW_GROWTH: yes
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
+        # TF_GPU_MEMORY_LIMIT: 4096
         OMP_NUM_THREADS: 1
         USE_CUDA: yes
       run: |
diff --git a/.github/workflows/self-scheduled.yml b/.github/workflows/self-scheduled.yml
index d963ab5920..0473949058 100644
--- a/.github/workflows/self-scheduled.yml
+++ b/.github/workflows/self-scheduled.yml
@@ -41,7 +41,7 @@ jobs:
 
     - name: Run all tests on GPU
       env:
-        TF_FORCE_GPU_ALLOW_GROWTH: yes
+        TF_FORCE_GPU_ALLOW_GROWTH: "true"
         OMP_NUM_THREADS: 1
         RUN_SLOW: yes
         USE_CUDA: yes
diff --git a/tests/test_modeling_tf_common.py b/tests/test_modeling_tf_common.py
index bcfb6bfe5d..e6f70d6bfa 100644
--- a/tests/test_modeling_tf_common.py
+++ b/tests/test_modeling_tf_common.py
@@ -21,14 +21,26 @@ import tempfile
 
 from transformers import is_tf_available, is_torch_available
 
-from .utils import require_tf
+from .utils import _tf_gpu_memory_limit, require_tf
 
 
 if is_tf_available():
     import tensorflow as tf
     import numpy as np
 
-    # from transformers.modeling_bert import BertModel, BertConfig, BERT_PRETRAINED_MODEL_ARCHIVE_MAP
+    if _tf_gpu_memory_limit is not None:
+        gpus = tf.config.list_physical_devices("GPU")
+        for gpu in gpus:
+            # Restrict TensorFlow to only allocate x GB of memory on the GPUs
+            try:
+                tf.config.experimental.set_virtual_device_configuration(
+                    gpu, [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
+                )
+                logical_gpus = tf.config.experimental.list_logical_devices("GPU")
+                print("Logical GPUs", logical_gpus)
+            except RuntimeError as e:
+                # Virtual devices must be set before GPUs have been initialized
+                print(e)
 
 
 def _config_zero_init(config):
diff --git a/tests/utils.py b/tests/utils.py
index 163628d3a7..6036cefd9d 100644
--- a/tests/utils.py
+++ b/tests/utils.py
@@ -29,8 +29,22 @@ def parse_flag_from_env(key, default=False):
     return _value
 
 
+def parse_int_from_env(key, default=None):
+    try:
+        value = os.environ[key]
+    except KeyError:
+        _value = default
+    else:
+        try:
+            _value = int(value)
+        except ValueError:
+            raise ValueError("If set, {} must be a int.".format(key))
+    return _value
+
+
 _run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
 _run_custom_tokenizers = parse_flag_from_env("RUN_CUSTOM_TOKENIZERS", default=False)
+_tf_gpu_memory_limit = parse_int_from_env("TF_GPU_MEMORY_LIMIT", default=None)
 
 
 def slow(test_case):
diff --git a/utils/link_tester.py b/utils/link_tester.py
index 0ef165c401..ff53eb7049 100644
--- a/utils/link_tester.py
+++ b/utils/link_tester.py
@@ -14,6 +14,9 @@ import requests
 REGEXP_FIND_S3_LINKS = r"""([\"'])(https:\/\/s3)(.*)?\1"""
 
 
+S3_BUCKET_PREFIX = "https://s3.amazonaws.com/models.huggingface.co/bert"
+
+
 def list_python_files_in_repository():
     """ List all python files in the repository.
 
@@ -36,7 +39,7 @@ def find_all_links(file_paths):
     for path in file_paths:
         links += scan_code_for_links(path)
 
-    return links
+    return [link for link in links if link != S3_BUCKET_PREFIX]
 
 
 def scan_code_for_links(source):