fix roll on CPU (#1840)

Author: lvyufeng

mindnlp/accelerate/big_modeling.py

@@ -73,13 +73,13 @@ def register_empty_parameter(module, name, param):
         if param is not None:
             kwargs = module._parameters[name].__dict__
             kwargs["requires_grad"] = param.requires_grad
-            module._parameters[name].assign_value(Tensor_(shape=(), dtype=module._parameters[name].dtype))
+            module._parameters[name].assign_value(Tensor_(shape=module._parameters[name].shape, dtype=module._parameters[name].dtype))
             module._parameters[name].meta = True
 
     def register_empty_buffer(module, name, buffer, persistent=True):
         old_register_buffer(module, name, buffer, persistent=persistent)
         if buffer is not None:
-            module._buffers[name].assign_value(Tensor_(shape=(), dtype=module._buffers[name].dtype))
+            module._buffers[name].assign_value(Tensor_(shape=module._parameters[name].shape, dtype=module._buffers[name].dtype))
             module._buffers[name].meta = True
 
     try:

mindnlp/core/ops/other.py

@@ -619,9 +619,12 @@ def repeat_interleave(input, repeats, dim=None):
     return input.repeat(repeats, dim)
 
 # roll
+DEVICE_TARGET = mindspore.get_context('device_target')
 def roll(input, shifts, dims=None):
     if use_pyboost():
         return mindspore.mint.roll(input, shifts, dims)
+    if DEVICE_TARGET == 'CPU':
+        return mindspore.numpy.roll(input, shifts, dims)
     return ops.roll(input, shifts, dims)
 
 # searchsorted

mindnlp/patch.py

@@ -2,6 +2,7 @@
 import numpy as np
 import mindspore
 from mindspore import Tensor
+from mindnlp.configs import GENERATOR_SEED
 
 def infer_value_for_BroadcastTo(x, shape):
     """Infer value for BroadcastTo op."""
@@ -18,4 +19,5 @@ def none_in_tuple_or_list(x):
     np_data = np.broadcast_to(x.asnumpy(), shape)
     return Tensor(np_data)
 
-mindspore.ops.operations.manually_defined.ops_def.infer_value_for_BroadcastTo = infer_value_for_BroadcastTo
+if GENERATOR_SEED:
+    mindspore.ops.operations.manually_defined.ops_def.infer_value_for_BroadcastTo = infer_value_for_BroadcastTo