Allow tests in examples to use cuda or fp16,if they are available (#5512)

* Allow tests in examples to use cuda or fp16,if they are available

The tests in examples didn't use the cuda or fp16 even if they where available.
- The text classification example (`run_glue.py`) didn't use the fp16 even if it was available but
  the device was take based on the availablity(cuda/cpu).
- The language-modeling example (`run_language_modeling.py`) was having `--no_cuda` argument
  which made the test to work without cuda. This example is having issue when running with fp16
  thus it not enabled (got an assertion error for perplexity due to it higher value).
- The cuda and fp16 is not enabled for question-answering example (`run_squad.py`) as it is having a
  difference in the f1 score.
- The text-generation example (`run_generation.py`) will take the cuda or fp16 whenever it is available.

Resolves some of: #5057

* Unwanted import of is_apex_available was removed

* Made changes to test examples file to have the pass --fp16 only if cuda and apex is avaliable
- run_glue.py: Removed the check for cuda and fp16.
- run_generation.py: Removed the check for cuda and fp16 also removed unwanted flag creation.

* Incorrectly sorted imports fixed

* The model needs to be converted to half precision

* Formatted single line if condition statement to multiline

* The torch_device also needed to be checked before running the test on examples
- The tests in examples which uses cuda should also depend from the USE_CUDA flag,
  similarly to the rest of the test suite. Even if we decide to set USE_CUDA to
  True by default, setting USE_CUDA to False should result in the examples not using CUDA

* Format some of the code in test_examples file

* The improper import of is_apex_available was sorted

* Formatted the code to keep the style standards

* The comma at the end of list giving a flake8 issue was fixed

* Import sort was fixed

* Removed the clean_test_dir function as its not used right now

examples/test_examples.py

@@ -22,7 +22,8 @@ from unittest.mock import patch
 
 import torch
 
-from transformers.testing_utils import TestCasePlus
+from transformers.file_utils import is_apex_available
+from transformers.testing_utils import TestCasePlus, torch_device
 
 
 SRC_DIRS = [
@@ -52,6 +53,11 @@ def get_setup_file():
     return args.f
 
 
+def is_cuda_and_apex_avaliable():
+    is_using_cuda = torch.cuda.is_available() and torch_device == "cuda"
+    return is_using_cuda and is_apex_available()
+
+
 class ExamplesTests(TestCasePlus):
     def test_run_glue(self):
         stream_handler = logging.StreamHandler(sys.stdout)
@@ -74,7 +80,13 @@ class ExamplesTests(TestCasePlus):
             --warmup_steps=2
             --seed=42
             --max_seq_length=128
-            """.split()
+            """
+        output_dir = "./tests/fixtures/tests_samples/temp_dir_{}".format(hash(testargs))
+        testargs += "--output_dir " + output_dir
+        testargs = testargs.split()
+
+        if is_cuda_and_apex_avaliable():
+            testargs.append("--fp16")
 
         with patch.object(sys, "argv", testargs):
             result = run_glue.main()
@@ -135,8 +147,13 @@ class ExamplesTests(TestCasePlus):
             --do_train
             --do_eval
             --num_train_epochs=1
-            --no_cuda
-            """.split()
+            """
+        output_dir = "./tests/fixtures/tests_samples/temp_dir_{}".format(hash(testargs))
+        testargs += "--output_dir " + output_dir
+        testargs = testargs.split()
+
+        if torch_device != "cuda":
+            testargs.append("--no_cuda")
 
         with patch.object(sys, "argv", testargs):
             result = run_language_modeling.main()
@@ -175,7 +192,14 @@ class ExamplesTests(TestCasePlus):
         logger.addHandler(stream_handler)
 
         testargs = ["run_generation.py", "--prompt=Hello", "--length=10", "--seed=42"]
-        model_type, model_name = ("--model_type=gpt2", "--model_name_or_path=sshleifer/tiny-gpt2")
+
+        if is_cuda_and_apex_avaliable():
+            testargs.append("--fp16")
+
+        model_type, model_name = (
+            "--model_type=gpt2",
+            "--model_name_or_path=sshleifer/tiny-gpt2",
+        )
         with patch.object(sys, "argv", testargs + [model_type, model_name]):
             result = run_generation.main()
             self.assertGreaterEqual(len(result[0]), 10)

examples/text-generation/run_generation.py

@@ -186,11 +186,20 @@ def main():
     parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
     parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
     parser.add_argument("--num_return_sequences", type=int, default=1, help="The number of samples to generate.")
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
     args = parser.parse_args()
 
     args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
     args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count()
 
+    logger.warning(
+        "device: %s, n_gpu: %s, 16-bits training: %s", args.device, args.n_gpu, args.fp16,
+    )
+
     set_seed(args)
 
     # Initialize the model and tokenizer
@@ -204,6 +213,9 @@ def main():
     model = model_class.from_pretrained(args.model_name_or_path)
     model.to(args.device)
 
+    if args.fp16:
+        model.half()
+
     args.length = adjust_length_to_model(args.length, max_sequence_length=model.config.max_position_embeddings)
     logger.info(args)