1 系统环境
硬件环境(Ascend/GPU/CPU): Ascend/GPU/CPU
MindSpore版本: mindspore=2.4
执行模式(PyNative/ Graph): Graph
Python版本: Python=3.9
操作系统平台: linux
2 报错信息
2.1 问题描述
使用mint.index_select在图模式如果没有unset RANK_TABLE_FILE,就会报错RuntimeError
2.2 报错信息
不开启unet:
=============================================================== FAILURES ===============================================================
__________________________________________________ test_index_select_forward_back[0] ___________________________________________________
mode = 0
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_index_select_forward_back(mode):
"""测试前向和反向传播,对比梯度"""
ms.set_context(mode=mode)
def forward_ms(x):
return mint.index_select(x, 0, ms_index).sum()
def forward_torch(x):
return torch.index_select(x, 0, torch_index).sum()
try:
input_data = [[1, 6, 2, 4], [7, 3, 8, 2], [2, 9, 11, 5]]
index = [0, 2]
ms_tensor, torch_tensor, ms_index, torch_index = create_tensors(input_data, ms.float32, torch.float32, index=index, requires_grad=True)
grad_fn_ms = value_and_grad(forward_ms)
output_ms, gradient_ms = grad_fn_ms(ms_tensor)
output_torch = forward_torch(torch_tensor)
output_torch.backward()
compare_results(output_ms, output_torch.detach())
compare_results(gradient_ms, torch_tensor.grad)
except Exception as e:
> raise e
test_index_select.py:167:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
test_index_select.py:160: in test_index_select_forward_back
output_ms, gradient_ms = grad_fn_ms(ms_tensor)
../anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py:960: in staging_specialize
out = _MindsporeFunctionExecutor(func, hash_obj, dyn_args, process_obj, jit_config)(*args, **kwargs)
../anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py:188: in wrapper
results = fn(*arg, **kwargs)
../anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py:582: in __call__
raise err
../anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py:579: in __call__
phase = self.compile(self.fn.__name__, *args_list, **kwargs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
self = <mindspore.common.api._MindsporeFunctionExecutor object at 0xffff8a646b80>, method_name = 'after_grad'
args = (Tensor(shape=[3, 4], dtype=Float32, value=
[[ 1.00000000e+00, 6.00000000e+00, 2.00000000e+00, 4.00000000e+00],
[ ...00000e+00, 8.00000000e+00, 2.00000000e+00],
[ 2.00000000e+00, 9.00000000e+00, 1.10000000e+01, 5.00000000e+00]]),)
kwargs = {}
compile_args = (Tensor(shape=[3, 4], dtype=Float32, value=
[[ 1.00000000e+00, 6.00000000e+00, 2.00000000e+00, 4.00000000e+00],
[ ...00000e+00, 8.00000000e+00, 2.00000000e+00],
[ 2.00000000e+00, 9.00000000e+00, 1.10000000e+01, 5.00000000e+00]]),)
key_id = '1876501395697841734756182555320576'
generate_name = 'mindspore.ops.composite.base.after_grad./home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/ops/composite/base.py.599.1734756182555320576'
echo_function_name = 'function "after_grad" at the file "/home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/ops/composite/base.py", line 599'
full_function_name = 'mindspore.ops.composite.base.after_grad./home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/ops/composite/base.py.599'
create_time = '1734756182555320576', key = 0, parameter_ids = ''
phase = 'mindspore.ops.composite.base.after_grad./home/ma-user/anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/ops/composite/base.py.599.1734756182555320576.0'
jit_config_dict = {'debug_level': 'RELEASE', 'exc_mode': 'auto', 'infer_boost': 'off', 'jit_level': '', ...}
def compile(self, method_name, *args, **kwargs):
"""Returns pipeline for the given args."""
# Check whether hook function registered on Cell object.
if self.obj and hasattr(self.obj, "_hook_fn_registered"):
if self.obj._hook_fn_registered():
logger.warning(f"For 'Cell', it's not support hook function when using 'jit' decorator. "
f"If you want to use hook function, please use context.set_context to set "
f"pynative mode and remove 'jit' decorator.")
# Chose dynamic shape tensors or actual input tensors as compile args.
compile_args = self._generate_compile_args(args)
key_id = self._get_key_id()
compile_args = get_auto_dynamic_shape_args_with_check_input_signature(compile_args, key_id,
self.input_signature)
# Restore the mutable attr for every arg.
compile_args = _restore_mutable_attr(args, compile_args)
self._compile_args = compile_args
generate_name, echo_function_name = self._get_generate_name()
# The full Function name
full_function_name = generate_name
create_time = ''
# Add key with obj
if self.obj is not None:
if self.obj.__module__ != self.fn.__module__:
logger.info(
f'The module of `self.obj`: `{self.obj.__module__}` is not same with the module of `self.fn`: '
f'`{self.fn.__module__}`')
self.obj.__parse_method__ = method_name
if isinstance(self.obj, ms.nn.Cell):
generate_name = generate_name + '.' + str(self.obj.create_time)
create_time = str(self.obj.create_time)
else:
generate_name = generate_name + '.' + str(self._create_time)
create_time = str(self._create_time)
generate_name = generate_name + '.' + str(id(self.obj))
full_function_name = generate_name
else:
# Different instance of same class may use same memory(means same obj_id) at diff times.
# To avoid unexpected phase matched, add create_time to generate_name.
generate_name = generate_name + '.' + str(self._create_time)
create_time = str(self._create_time)
self.enable_tuple_broaden = False
if hasattr(self.obj, "enable_tuple_broaden"):
self.enable_tuple_broaden = self.obj.enable_tuple_broaden
self._graph_executor.set_enable_tuple_broaden(self.enable_tuple_broaden)
key = self._graph_executor.generate_arguments_key(self.fn, compile_args, kwargs, self.enable_tuple_broaden)
parameter_ids = _get_parameter_ids(args, kwargs)
if parameter_ids != "":
key = str(key) + '.' + parameter_ids
phase = generate_name + '.' + str(key)
update_auto_dynamic_shape_phase_with_check_input_signature(compile_args, key_id, phase, self.input_signature)
if phase in ms_compile_cache:
# Release resource should be released when CompileInner won't be executed, such as cur_convert_input_
# generated in generate_arguments_key.
self._graph_executor.clear_compile_arguments_resource()
return phase
_check_recompile(self.obj, compile_args, kwargs, full_function_name, create_time, echo_function_name)
# If enable compile cache, get the dependency files list and set to graph executor.
self._set_compile_cache_dep_files()
if self.jit_config_dict:
self._graph_executor.set_jit_config(self.jit_config_dict)
else:
jit_config_dict = JitConfig().jit_config_dict
self._graph_executor.set_jit_config(jit_config_dict)
if self.obj is None:
# Set an attribute to fn as an identifier.
if isinstance(self.fn, types.MethodType):
setattr(self.fn.__func__, "__jit_function__", True)
else:
setattr(self.fn, "__jit_function__", True)
> is_compile = self._graph_executor.compile(self.fn, compile_args, kwargs, phase, True)
E RuntimeError: Compile graph kernel_graph0 failed.
E
E ----------------------------------------------------
E - C++ Call Stack: (For framework developers)
E ----------------------------------------------------
E mindspore/ccsrc/plugin/device/ascend/hal/hardware/ge_graph_executor.cc:636 CompileGraph
../anaconda3/envs/MindSpore/lib/python3.9/site-packages/mindspore/common/api.py:674: RuntimeError复制
开启:
test_index_select.py:165: in test_index_select_forward_back
compare_results(gradient_ms, torch_tensor.grad)_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
ms_result = Tensor(shape=[3, 4], dtype=Float32, value=
[[-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00],
[-7...1860191e-01, 6.35216951e-01, 2.55284572e+00],
[-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00]])
torch_result = tensor([[1., 1., 1., 1.],
[0., 0., 0., 0.],
[1., 1., 1., 1.]]), atol = 0.001
def compare_results(ms_result, torch_result, atol=1e-3):
> assert np.allclose(ms_result.asnumpy(), torch_result.numpy(), atol=atol)
E assert False
E + where False = <function allclose at 0xffff078191f0>(array([[-0.09153102, 1.2882982 , -1.1313928 , -2.2087781 ],\n [-0.75677174, -0.4818602 , 0.63521695, 2.5528457 ],\n [-0.09153102, 1.2882982 , -1.1313928 , -2.2087781 ]],\n dtype=float32), array([[1., 1., 1., 1.],\n [0., 0., 0., 0.],\n [1., 1., 1., 1.]], dtype=float32), atol=0.001)
E + where <function allclose at 0xffff078191f0> = np.allclose
E + and array([[-0.09153102, 1.2882982 , -1.1313928 , -2.2087781 ],\n [-0.75677174, -0.4818602 , 0.63521695, 2.5528457 ],\n [-0.09153102, 1.2882982 , -1.1313928 , -2.2087781 ]],\n dtype=float32) = <bound method Tensor.asnumpy of Tensor(shape=[3, 4], dtype=Float32, value=\n[[-9.15310234e-02, 1.28829825e+00, -1.1313...860191e-01, 6.35216951e-01, 2.55284572e+00],\n [-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00]])>()
E + where <bound method Tensor.asnumpy of Tensor(shape=[3, 4], dtype=Float32, value=\n[[-9.15310234e-02, 1.28829825e+00, -1.1313...860191e-01, 6.35216951e-01, 2.55284572e+00],\n [-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00]])> = Tensor(shape=[3, 4], dtype=Float32, value=\n[[-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00],\n [-7...1860191e-01, 6.35216951e-01, 2.55284572e+00],\n [-9.15310234e-02, 1.28829825e+00, -1.13139284e+00, -2.20877814e+00]]).asnumpy
E + and array([[1., 1., 1., 1.],\n [0., 0., 0., 0.],\n [1., 1., 1., 1.]], dtype=float32) = <built-in method numpy of Tensor object at 0xffffa45d3040>()
E + where <built-in method numpy of Tensor object at 0xffffa45d3040> = tensor([[1., 1., 1., 1.],\n [0., 0., 0., 0.],\n [1., 1., 1., 1.]]).numpy
test_index_select.py:28: AssertionError复制
2.3 脚本信息
import pytest
import numpy as np
import mindspore as ms
from mindspore import Tensor, value_and_grad, mint
import torch
def create_tensors(input_data, ms_dtype, torch_dtype, index=None, requires_grad=False):
ms_tensor = Tensor(input_data, ms_dtype)
torch_tensor = torch.tensor(input_data, dtype=torch_dtype, requires_grad=requires_grad)
if index is not None:
ms_index = Tensor(index, ms.int32)
torch_index = torch.tensor(index, dtype=torch.long)
else:
ms_index = None
torch_index = None
return ms_tensor, torch_tensor, ms_index, torch_index
def perform_index_select(ms_tensor, torch_tensor, dim, ms_index, torch_index):
ms_result = mint.index_select(ms_tensor, dim, ms_index)
torch_result = torch.index_select(torch_tensor, dim, torch_index)
return ms_result, torch_result
def compare_results(ms_result, torch_result, atol=1e-3):
assert np.allclose(ms_result.asnumpy(), torch_result.numpy(), atol=atol)
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_index_select_forward_back(mode):
"""测试前向和反向传播,对比梯度"""
ms.set_context(mode=mode)
def forward_ms(x):
return mint.index_select(x, 0, ms_index).sum()
def forward_torch(x):
return torch.index_select(x, 0, torch_index).sum()
try:
input_data = [[1, 6, 2, 4], [7, 3, 8, 2], [2, 9, 11, 5]]
index = [0, 2]
ms_tensor, torch_tensor, ms_index, torch_index = create_tensors(input_data, ms.float32, torch.float32, index=index, requires_grad=True)
grad_fn_ms = value_and_grad(forward_ms)
output_ms, gradient_ms = grad_fn_ms(ms_tensor)
output_torch = forward_torch(torch_tensor)
output_torch.backward()
compare_results(output_ms, output_torch.detach())
compare_results(gradient_ms, torch_tensor.grad)
except Exception as e:
raise e
在终端分别使用或者不使用unset RANK_TABLE_FILE,再使用下面的脚本。
3 根因分析
根因:mint接口在图模式下暂不支持图下沉模式。
查看mint接口的相关说明,里面有提及到mint接口的限制。
mindspore.mint提供了大量的functional、nn、优化器接口,API用法及功能等与业界主流用法一致,方便用户参考使用。 mint接口当前是实验性接口,在图编译模式为O0和PyNative模式下性能比ops更优。当前暂不支持图下沉模式及CPU、GPU后端,后续会逐步完善。
然后再参看set_context的接口说明,里面有图模式的相关设置的说明。Altas训练产品为O2,其余产品均为O0
jit_config (dict) - 设置全局编译选项的配置,只在使用Cell或者jit装饰器定义的网络中生效,默认不设置。 context设置全局jit config,而JitConfig设置局部网络的jit config,二者同时存在时,全局jit config不会覆盖局部网络的jit config。
jit_level (str): 用来控制编译优化级别。默认值为空,框架根据产品类别自动选择优化级别,Altas训练产品为O2,其余产品均为O0。动态Shape场景须为O0/O1,不支持O2。其值范围如下:
O0: 除必要影响功能的优化外,其他优化均关闭,使用逐算子执行的执行方式。
O1: 使能常用优化和自动算子融合优化,使用逐算子执行的执行方式。
O2: 开启极致性能优化,使用下沉的执行方式。
4 解决方案
因为mint接口不支持图下沉模式,所以代码手动设置 jit_level为O0
ms.set_context(jit_level=“O0”),修改后代码如下
import pytest
import numpy as np
import mindspore as ms
from mindspore import Tensor, value_and_grad, mint
import torch
def create_tensors(input_data, ms_dtype, torch_dtype, index=None, requires_grad=False):
ms_tensor = Tensor(input_data, ms_dtype)
torch_tensor = torch.tensor(input_data, dtype=torch_dtype, requires_grad=requires_grad)
if index is not None:
ms_index = Tensor(index, ms.int32)
torch_index = torch.tensor(index, dtype=torch.long)
else:
ms_index = None
torch_index = None
return ms_tensor, torch_tensor, ms_index, torch_index
def perform_index_select(ms_tensor, torch_tensor, dim, ms_index, torch_index):
ms_result = mint.index_select(ms_tensor, dim, ms_index)
torch_result = torch.index_select(torch_tensor, dim, torch_index)
return ms_result, torch_result
def compare_results(ms_result, torch_result, atol=1e-3):
assert np.allclose(ms_result.asnumpy(), torch_result.numpy(), atol=atol)
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_index_select_forward_back(mode):
"""测试前向和反向传播,对比梯度"""
ms.set_context(mode=mode)
ms.set_context(jit_level="O0")
def forward_ms(x):
return mint.index_select(x, 0, ms_index).sum()
def forward_torch(x):
return torch.index_select(x, 0, torch_index).sum()
try:
input_data = [[1, 6, 2, 4], [7, 3, 8, 2], [2, 9, 11, 5]]
index = [0, 2]
ms_tensor, torch_tensor, ms_index, torch_index = create_tensors(input_data, ms.float32, torch.float32, index=index, requires_grad=True)
grad_fn_ms = value_and_grad(forward_ms)
output_ms, gradient_ms = grad_fn_ms(ms_tensor)
output_torch = forward_torch(torch_tensor)
output_torch.backward()
compare_results(output_ms, output_torch.detach())
compare_results(gradient_ms, torch_tensor.grad)
except Exception as e:
raise e