fix roll not support on CPU (#932)

Author: lvyufeng

llm/inference/bark/inference.py

@@ -2,7 +2,6 @@
 import mindspore
 from IPython.display import Audio
 from mindnlp.transformers.models.bark import BarkModel, BarkProcessor
-mindspore.set_context(pynative_synchronize=True, device_target="CPU")
 
 
 voice_preset = None

tests/ut/transformers/models/bark/test_modeling_bark.py

@@ -58,7 +58,6 @@
         BarkSemanticModel,
     )
 
-# mindspore.set_context(pynative_synchronize=True)
 
 class BarkSemanticModelTester:
     def __init__(

tests/ut/transformers/models/big_bird/test_modeling_big_bird.py

@@ -49,7 +49,6 @@
         BIG_BIRD_PRETRAINED_MODEL_ARCHIVE_LIST,
     )
 
-# mindspore.set_context(pynative_synchronize=True)
 
 class BigBirdModelTester:
     def __init__(

tests/ut/transformers/models/chatglm/test_modeling_graph_chatglm.py

@@ -9,7 +9,6 @@
 
 from mindspore._c_expression import _framework_profiler_step_start
 from mindspore._c_expression import _framework_profiler_step_end
-# mindspore.set_context(pynative_synchronize=True)
 # mindspore.set_context(mode=mindspore.GRAPH_MODE, jit_syntax_level=mindspore.LAX, save_graphs=True, save_graphs_path="./graph")
 
 def set_random_seed(seed):

tests/ut/transformers/models/mixtral/test_modeling_mixtral.py

@@ -37,7 +37,6 @@
 
     from mindnlp.transformers import MixtralForCausalLM, MixtralForSequenceClassification, MixtralModel
 
-# mindspore.set_context(pynative_synchronize=True)
 
 class MixtralModelTester:
     # Copied from tests.models.mistral.test_modeling_mistral.MistralModelTester.__init__

tests/ut/transformers/models/reformer/test_modeling_reformer.py

@@ -45,7 +45,6 @@
         ReformerTokenizer,
     )
 
-mindspore.set_context(pynative_synchronize=True)
 
 class ReformerModelTester:
     def __init__(
@@ -275,8 +274,8 @@ def create_and_check_reformer_model_with_attn_mask(
             [half_input_ids, ids_tensor((self.batch_size, half_seq_len), self.vocab_size)],
             axis=-1,
         )
-        input_ids_roll = ops.roll(input_ids_roll, roll, dims=-1)
-        attn_mask_roll = ops.roll(attn_mask, roll, dims=-1)
+        input_ids_roll = mindspore.numpy.roll(input_ids_roll, roll, axis=-1)
+        attn_mask_roll = mindspore.numpy.roll(attn_mask, roll, axis=-1)
 
         output_padded = model(input_ids_padded, attention_mask=attn_mask)[0][:, :half_seq_len]
         output_padded_rolled = model(input_ids_roll, attention_mask=attn_mask_roll)[0][:, roll : half_seq_len + roll]