Remove setting of PT_HPU_LAZY_MODE=2 in training_args.py (huggingface…

…#625)

docs/Dockerfile

@@ -16,6 +16,6 @@ RUN npm install npm@9.8.1 -g && \
     npm install n -g && \
     n latest
 
-RUN git clone $clone_url && cd optimum-habana && git checkout $commit_sha
+RUN git clone $clone_url optimum-habana && cd optimum-habana && git checkout $commit_sha
 RUN python3 -m pip install --no-cache-dir --upgrade pip
 RUN python3 -m pip install --no-cache-dir ./optimum-habana[quality]

optimum/habana/transformers/training_args.py

@@ -648,12 +648,11 @@ def _setup_devices(self) -> "torch.device":
 
             if self.use_lazy_mode:
                 logger.info("Enabled lazy mode.")
-            # TODO: remove the block below when upgrade to SynapseAI 1.13 is done
-            # as eager mode will not be available anymore
             elif not self.torch_compile:
                 if os.getenv("PT_HPU_LAZY_MODE", "1") != "0":
-                    os.environ["PT_HPU_LAZY_MODE"] = "2"
-                logger.info("Enabled eager mode because use_lazy_mode=False.")
+                    raise ValueError(
+                        "Lazy mode or compile mode not enabled => eager mode should be enabled using PT_HPU_LAZY_MODE=0"
+                    )
 
             if self.deepspeed:
                 # Need to do similar for Accelerator init

tests/test_trainer.py

@@ -29,6 +29,7 @@
 import numpy as np
 from huggingface_hub import HfFolder, delete_repo, list_repo_commits
 from parameterized import parameterized
+from pytest import mark
 from requests.exceptions import HTTPError
 from transformers import IntervalStrategy, PretrainedConfig, is_torch_available
 from transformers.hyperparameter_search import ALL_HYPERPARAMETER_SEARCH_BACKENDS
@@ -706,6 +707,7 @@ def setUp(self):
         self.n_epochs = args.num_train_epochs
         self.batch_size = args.train_batch_size
 
+    @mark.skip("Skip this test until PT_HPU_LAZY_MODE=0 is set as default for all tests")
     def test_eager_mode(self):
         train_dataset = RegressionDataset()
         eval_dataset = RegressionDataset()