[SOT][CINN] Only create SymbolicVariable when value >= 2 & add corresponding guards and C++ constraints

[GoldenStain]

PR Category

Execute Infrastructure

PR Types

Performance

Description

1. Conditional symbolic variable creation in SOT

Modification: Symbolic variables are now only instantiated when tensor dimensions ≥ 2 in the SOT compiler phase.

Impact: Reduce the number of kernels generated by CINN in later pass.

2. Add corresponding Guards

Current approach: Implemented temporary validation using StringifiedGuard
Future plan: Replace it with FasterGuard.

3. Cross-language constraint (Python → C++)

Implementation:

1. Introduced ConstrainedInputSpec subclass extending InputSpec
2. Designed constraint storage format:

   # ConstrainedInputSpec.ranges: [list[tuple[int, int | None, int | None]]]
   [
     (index of dim, min, max)
   ]

3. Established serialization pipeline: Python list → PyBind11 type conversion → C++ InputDynamicDimSpec
   Extract the constraint information from ConstrainedInputSpec, then pass it to ConcreteProgram, where it is finally passed on to PartialProgramLayer, where we bind the constraint information using interface exposed by Pybind.

4. Add new "Append" mode for SpecifyInputDynamicDim as AppendInputDynamicDim

This new interface can be useful in future expansion.

5. Add new __init__() arguments for PartialProgramLayer and ConcreteProgram

Now both the 2 classes have a new key-word argument named constraints, whose default value is None.

6. Add new __init__() argument for MetaInfo

Add a new positional argument spec_name for MetaInfo in __init__(), remember to pass it when creating new MetaInfo.

PCard-66972

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[SigureMo]

其他的看起来没啥问题

[SigureMo]

突然想起来这个 PR 是不是没把 >=2 加到 guard 里？从正确性上来讲这是不够的，需要测一下：

shape = [2, 2, 3] # cache miss，静态 shape
shape = [3, 2, 3] # cache miss，动态 shape
shape = [1, 2, 3] # cache miss，静态 shape，按我理解现在的实现这里是会 hit 到上面的动态 shape 的，这是不对的

[GoldenStain]

突然想起来这个 PR 是不是没把 >=2 加到 guard 里？从正确性上来讲这是不够的，需要测一下：

shape = [2, 2, 3] # cache miss，静态 shape
shape = [3, 2, 3] # cache miss，动态 shape
shape = [1, 2, 3] # cache miss，静态 shape，按我理解现在的实现这里是会 hit 到上面的动态 shape 的，这是不对的

guard的修改在TODO的第三项，在完成了第三个改动之后再加入相应的单测。

[GoldenStain]

向PartialProgramLayer添加参数时：

def __init__(
        self,
        main_program,
        inputs,
        outputs,
        constrained_inputs=None,
        parameters=None,
        **kwargs,
    ):

把constrained_inputs放在了parameters的前面，导致按位置传递参数时，拿不到正确的parameters，导致scope拿不到正确的信息，在save后load时挂掉。

[SigureMo]

其余目测没啥大问题

[SigureMo]

另外 approval 也需要看一下，看起来可能是信息太短之类的问题

[zyfncg]

可以写成两个函数接口，语义上更清晰

[zyfncg]

可以显式设置个min max的默认值

[GoldenStain]

Range结构体内部有min max的默认值

[SigureMo]

Great work!

LGTMeow