[cuda ext tests] fixing tests (#11619)

* fixing tests

* cleanup

.github/workflows/self-scheduled.yml

@@ -261,6 +261,7 @@ jobs:
 
       - name: Install dependencies
         run: |
+          apt -y update && apt install -y libaio-dev
           pip install --upgrade pip
           pip install .[testing,deepspeed]
 
@@ -301,6 +302,7 @@ jobs:
 
       - name: Install dependencies
         run: |
+          apt -y update && apt install -y libaio-dev
           pip install --upgrade pip
           pip install .[testing,deepspeed,fairscale]
 

tests/deepspeed/test_deepspeed.py

@@ -318,9 +318,10 @@ class TrainerIntegrationDeepSpeed(TestCasePlus, TrainerIntegrationCommon):
             yes_grad_accum_b = yes_grad_accum_trainer.model.b.item()
             self.assertNotEqual(yes_grad_accum_a, a)
 
-        # training with half the batch size but accumulation steps as 2 should give the same weights
-        self.assertEqual(no_grad_accum_a, yes_grad_accum_a)
-        self.assertEqual(no_grad_accum_b, yes_grad_accum_b)
+        # training with half the batch size but accumulation steps as 2 should give the same
+        # weights, but sometimes get a slight difference still of 1e-6
+        self.assertAlmostEqual(no_grad_accum_a, yes_grad_accum_a, places=5)
+        self.assertAlmostEqual(no_grad_accum_b, yes_grad_accum_b, places=5)
 
         # see the note above how to get identical loss on a small bs
         self.assertAlmostEqual(no_grad_accum_loss, yes_grad_accum_loss, places=5)

tests/extended/test_trainer_ext.py

@@ -167,8 +167,8 @@ class TestTrainerExt(TestCasePlus):
         # test if do_predict saves generations and metrics
         contents = os.listdir(output_dir)
         contents = {os.path.basename(p) for p in contents}
-        assert "test_generations.txt" in contents
-        assert "test_results.json" in contents
+        assert "generated_predictions.txt" in contents
+        assert "predict_results.json" in contents
 
     def run_trainer(
         self,