torch.动态值

constrain_as_size_example

注意

标签: torch.dynamic-value, torch.escape-hatch

支持级别：支持

原始源代码：

import torch

结果:

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, arg0_1: "i64[]"):
                _local_scalar_dense: "Sym(u4)" = torch.ops.aten._local_scalar_dense.default(arg0_1);  arg0_1 = None

                ge: "Sym(u4 >= 0)" = _local_scalar_dense >= 0
            scalar_tensor: "f32[]" = torch.ops.aten.scalar_tensor.default(ge);  ge = None
            _assert_async = torch.ops.aten._assert_async.msg(scalar_tensor, '_local_scalar_dense 超出了内联约束 [0, 5]。');  scalar_tensor = None
            le: "Sym(u4 <= 5)" = _local_scalar_dense <= 5
            scalar_tensor_1: "f32[]" = torch.ops.aten.scalar_tensor.default(le);  le = None
            _assert_async_1 = torch.ops.aten._assert_async.msg(scalar_tensor_1, '_local_scalar_dense 超出了内联约束 [0, 5]。');  scalar_tensor_1 = None

                sym_constrain_range_for_size = torch.ops.aten.sym_constrain_range_for_size.default(_local_scalar_dense, min = 0, max = 5)

                ones: "f32[u4, 5]" = torch.ops.aten.ones.default([_local_scalar_dense, 5], device = device(type='cpu'), pin_memory = False);  _local_scalar_dense = None
            return (ones,)

Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='arg0_1'), target=None, persistent=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='ones'), target=None)])
Range constraints: {u0: ValueRanges(lower=0, upper=5, is_bool=False), u1: ValueRanges(lower=0, upper=5, is_bool=False), u4: ValueRanges(lower=0, upper=5, is_bool=False)}

constrain_as_value_example

注意

标签: torch.dynamic-value, torch.escape-hatch

支持级别：支持

原始源代码：

import torch

结果:

ExportedProgram:
    class GraphModule(torch.nn.Module):
        def forward(self, arg0_1: "i64[]", arg1_1: "f32[5, 5]"):
                _local_scalar_dense: "Sym(u4)" = torch.ops.aten._local_scalar_dense.default(arg0_1);  arg0_1 = None

                ge: "Sym(u4 >= 0)" = _local_scalar_dense >= 0
            scalar_tensor: "f32[]" = torch.ops.aten.scalar_tensor.default(ge);  ge = None
            _assert_async = torch.ops.aten._assert_async.msg(scalar_tensor, '_local_scalar_dense 超出了内联约束 [0, 5]。');  scalar_tensor = None
            le: "Sym(u4 <= 5)" = _local_scalar_dense <= 5
            scalar_tensor_1: "f32[]" = torch.ops.aten.scalar_tensor.default(le);  le = None
            _assert_async_1 = torch.ops.aten._assert_async.msg(scalar_tensor_1, '_local_scalar_dense 超出了内联约束 [0, 5]。');  scalar_tensor_1 = None

                sym_constrain_range = torch.ops.aten.sym_constrain_range.default(_local_scalar_dense, min = 0, max = 5);  _local_scalar_dense = None

                sin: "f32[5, 5]" = torch.ops.aten.sin.default(arg1_1);  arg1_1 = None
            return (sin,)

Graph signature: ExportGraphSignature(input_specs=[InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='arg0_1'), target=None, persistent=None), InputSpec(kind=<InputKind.USER_INPUT: 1>, arg=TensorArgument(name='arg1_1'), target=None, persistent=None)], output_specs=[OutputSpec(kind=<OutputKind.USER_OUTPUT: 1>, arg=TensorArgument(name='sin'), target=None)])
Range constraints: {u0: ValueRanges(lower=0, upper=5, is_bool=False), u1: ValueRanges(lower=0, upper=5, is_bool=False), u4: ValueRanges(lower=0, upper=5, is_bool=False)}