support multiprocess inference on 910A (#2124)

Author: lvyufeng

mindnlp/__init__.py

@@ -36,6 +36,8 @@
 except:
     disable_multi_thread = None
 # for different ascend devices
+    context.set_context(device_target='CPU')
+
 if platform.system().lower() == 'linux':
     SOC = MSContext.get_instance().get_ascend_soc_version()
     if ('910b' not in SOC and '310' not in SOC) or version.parse(mindspore.__version__) < version.parse('2.4.0'):

mindnlp/accelerate/__init__.py

@@ -1,18 +0,0 @@
-import sys
-import accelerate
-from transformers.utils import _LazyModule
-
-_import_structure = {
-    "utils": [
-        'DistributedType',
-
-    ]
-}
-
-sys.modules[__name__] = _LazyModule(
-    'accelerate',
-    accelerate.__file__,
-    _import_structure,
-    module_spec=__spec__,
-    extra_objects={"__version__": accelerate.__version__},
-)

mindnlp/core/ops/fft_op.py renamed to mindnlp/accelerate/utils/__init__.py

@@ -6,6 +6,8 @@
 
 DEFAULT_DTYPE, DEFAULT_DEVICE = float32, device_('cpu')
 
+DEVICE_IN_CONTEXT = None
+
 AUTO_CAST_DTYE = {
     'cuda': float16,
     'cpu': bfloat16,
@@ -41,6 +43,16 @@ def get_default_device():
     """get default dtype"""
     return DEFAULT_DEVICE
 
+def set_device_in_context(device):
+    global DEVICE_IN_CONTEXT
+    DEVICE_IN_CONTEXT = device
+
+def get_device_in_context():
+    """get default dtype"""
+    if DEVICE_IN_CONTEXT is None:
+        return get_default_device()
+    return DEVICE_IN_CONTEXT
+
 bits_map = {
 
 }

mindnlp/accelerate/utils/modeling.py

@@ -0,0 +1,159 @@
+import numbers
+from mindspore import ops
+from mindspore.ops.auto_generate import gen_ops_prim
+from mindspore.ops.auto_generate import pyboost_inner_prim
+from mindspore._c_expression import _empty_instance
+
+from mindnlp import core
+from mindnlp.core._C import default_generator
+
+op_list = list(filter(lambda s: s.endswith("_op"), dir(gen_ops_prim)))
+
+__all__ = []
+
+for op_name in op_list:
+    func_name = op_name.replace('_op', '')
+    __all__.append(func_name)
+    globals()[func_name] = getattr(gen_ops_prim, op_name).__class__().set_device('Ascend')
+
+def empty(*args, **kwargs):
+    return _empty_instance(*args, **kwargs, device='Ascend')
+
+def reduce_any(input, dim, keepdim):
+    if dim is None:
+        dim = ()
+    return pyboost_inner_prim.reduce_any_impl(input, dim, keepdim)
+
+__all__.append('reduce_any')
+
+def reduce_all(input, dim, keepdim):
+    if dim is None:
+        dim = ()
+    return pyboost_inner_prim.reduce_all_impl(input, dim, keepdim)
+
+__all__.append('reduce_all')
+
+broadcast_to_op = ops.Primitive('BroadcastTo').set_device('Ascend')
+def broadcast_to(*args):
+    return broadcast_to_op(*args)
+
+__all__.append('broadcast_to')
+
+cast_op = ops.Cast().set_device('Ascend')
+def cast(*args):
+    return cast_op(*args)
+
+__all__.append('cast')
+
+zeros_op = ops.Zeros().set_device('Ascend')
+def zeros(*args):
+    return zeros_op(*args)
+
+__all__.append('zeros')
+
+def softmax(*args):
+    return pyboost_inner_prim.softmax_impl(*args)
+
+__all__.append('softmax')
+
+def dropout_ext(input, p):
+    seed, offset = default_generator._step(12)  # pylint: disable=protected-access
+    return gen_ops_prim.dropout_ext_op(input, p, seed, offset)
+
+def squeeze(*args):
+    return pyboost_inner_prim.squeeze_impl(*args)
+
+__all__.append('squeeze')
+
+ones_op = ops.Ones().set_device('Ascend')
+def ones(*args):
+    return ones_op(*args)
+
+__all__.append('ones')
+
+def nllloss(*args):
+    return pyboost_inner_prim.nllloss_impl(*args)
+
+__all__.append('nllloss')
+
+def repeat_elements(*args):
+    return ops.repeat_elements(*args)
+
+__all__.append('repeat_elements')
+
+def concat(*args):
+    return pyboost_inner_prim.concat_impl(*args)
+
+__all__.append('concat')
+
+def multinomial_ext(input, num_samples, replacement, generator):
+    seed, offset = generator._step(12)  # pylint: disable=protected-access
+    return gen_ops_prim.multinomial_ext_op(input, num_samples, replacement, seed, offset)
+
+def isclose(*args):
+    return pyboost_inner_prim.isclose_impl(*args)
+
+__all__.append('isclose')
+
+tile_op = ops.Primitive('Tile').set_device('Ascend')
+def tile(*args):
+    return tile_op(*args)
+
+__all__.append('tile')
+
+def pad_v3(input_x, padding, mode='constant', value=None):
+    pad_op = ops.PadV3(mode=mode, paddings_contiguous=True).set_device('CPU')
+    if isinstance(value, (float, int)):
+        value = core.tensor(value, dtype=input_x.dtype)
+    return pad_op(input_x, padding, value)
+
+__all__.append('pad_v3')
+
+def inplace_uniform(input, from_, to_, generator_):
+    seed, offset = generator_._step(12)
+    return gen_ops_prim.inplace_uniform_op(input, from_, to_, seed, offset)
+
+def binary_cross_entropy_with_logits(*args):
+    return pyboost_inner_prim.binary_cross_entropy_with_logits_impl(*args)
+
+__all__.append('binary_cross_entropy_with_logits')
+
+def gather(input_params, input_indices, axis, batch_dims=0):
+    return ops.gather(input_params, input_indices, axis, batch_dims)
+
+__all__.append('gather')
+
+def randint(low, high, shape, dtype, generator):
+    seed, offset = generator._step(12)  # pylint: disable=protected-access
+    return gen_ops_prim.randint_op(low, high, shape, seed, offset, dtype)
+
+def stack_ext(*args):
+    return pyboost_inner_prim.stack_ext_impl(*args)
+
+__all__.append('stack_ext')
+
+def argmax_with_value(*args):
+    return pyboost_inner_prim.argmax_with_value_impl(*args)
+
+__all__.append('argmax_with_value')
+
+right_shift_op = ops.RightShift().set_device('Ascend')
+def right_shift(input, other):
+    if isinstance(other, numbers.Number):
+        other = core.Tensor(other, input.dtype)
+    return right_shift_op(input, other)
+
+tensor_mul = ops.Mul().set_device('Ascend')
+tensor_pow = ops.Pow().set_device('Ascend')
+def ldexp(input, other):
+    out = tensor_mul(input, tensor_pow(2.0, other))
+    return out
+
+__all__.append('ldexp')
+
+def reverse_v2(input, dims):
+    if isinstance(dims, int):
+        dims = (dims,)
+    return pyboost_inner_prim.reverse_v2_impl(input, dims)
+
+__all__.append('reverse_v2')