| [feature] inductor_npu_ext: stride/shape check for fused kernel tools
Co-authored-by: yurongkun<yurongkun@huawei.com>
# message auto-generated for no-merge-commit merge:
!2945 merge inductor_check_0421 into master
[feature] inductor_npu_ext: stride/shape check for fused kernel tools
Created-by: yurongkun
Commit-by: yurongkun
Merged-by: ascend-robot
Description: stride/shape check for fused kernel tools:
indcutor_npu_ext新增融合kernel校验工具。不同于inductor原生assert_size功能,本功能主要针对融合kernel
实现功能如下:
1)新增TORCHINDUCTOR_NPU_EXT_LAYOUT_CHECK用于决定是否直接error报错,默认为关
2)打开原生TORCH_COMPILE_DEBUG或者TORCHINDUCTOR_NPU_EXT_LAYOUT_CHECK会生成对比代码.TORCH_COMPILE_DEBUG=1时仅打印日志warning级别
3)shape dim<=1不会校验stride信息,其他场景正常校验
4)针对SymInt/sympy.Expr/int三种场景的输入信息,做对应适配。_safe_int_convert
5)在output.py文件中新增maybe_check_fused_input_layout的codegen代码并在kernel call之前执行
6)新增UT 验证动态和静态图场景是否符合预期
See merge request: Ascend/torchair!2945 | 15 天前 |