Merge branch 'main' into reward-refactor

Author: qgallouedec

tests/slow/test_grpo_slow.py

@@ -33,7 +33,6 @@
 )
 from transformers.testing_utils import (
     backend_empty_cache,
-    require_bitsandbytes,
     require_flash_attn,
     require_liger_kernel,
     require_peft,
@@ -45,7 +44,7 @@
 from trl import GRPOConfig, GRPOTrainer
 from trl.trainer.utils import get_kbit_device_map
 
-from ..testing_utils import TrlTestCase, require_vllm
+from ..testing_utils import TrlTestCase, require_bitsandbytes, require_vllm
 from .testing_constants import MODELS_TO_TEST
 
 

tests/test_sft_trainer.py

@@ -20,19 +20,13 @@
 from datasets import load_dataset
 from parameterized import parameterized
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from transformers.testing_utils import (
-    require_bitsandbytes,
-    require_flash_attn,
-    require_liger_kernel,
-    require_peft,
-    require_vision,
-)
+from transformers.testing_utils import require_flash_attn, require_liger_kernel, require_peft, require_vision
 from transformers.utils import is_peft_available
 
 from trl import SFTConfig, SFTTrainer
 from trl.trainer.sft_trainer import DataCollatorForLanguageModeling, dft_loss
 
-from .testing_utils import TrlTestCase, ignore_warnings
+from .testing_utils import TrlTestCase, ignore_warnings, require_bitsandbytes
 
 
 if is_peft_available():