【Prim】support higher order autodiff for dy2static+composite #53171
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
❌ The PR is not created using PR's template. You can refer to this Demo. |
cxxly
force-pushed
the
prim_higher_order_autodiff
branch
from
883bcd2
to
ab25750
Compare
cxxly
changed the title
|
Sorry to inform you that ab25750's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
cxxly
force-pushed
the
prim_higher_order_autodiff
branch
3 times, most recently
from
94120d9
to
aee095e
Compare
cxxly
force-pushed
the
prim_higher_order_autodiff
branch
from
aee095e
to
40fd316
Compare
cxxly
force-pushed
the
prim_higher_order_autodiff
branch
from
40fd316
to
aa74cfe
Compare
| ): | ||
| composite_block.create_var(name=name) | ||
| op.desc.infer_var_type(composite_block.desc) | ||
| op.desc.infer_shape(composite_block.desc) |
There was a problem hiding this comment.
当前Block只包含前向OP列表，只需要对于没有输出的OP创建输出并推断形状，不需要将OP重复添加到block中。infershape_for_composite 会重复创建OP，比如fill_any_like
| # this patch for primitive operators. | ||
| for arg in grad_op_desc.output_arg_names(): | ||
| if arg in new_vars: | ||
| _infer_var_data_type_shape_(arg, block) |
There was a problem hiding this comment.
这部分修改是否可以合并至1784 else 分支中
There was a problem hiding this comment.
可以的，只不过这样代码是一个非常Hack的补丁(Paddle 1.x版本已存在)，单独放一段逻辑更容易引起注意
PR types
New features
PR changes
Others
Description
Pcard-66975
动转静+组合模式适配高阶微分部分逻辑：
_strip_grad_suffix_适配，高阶微分模式下， 会产生z@GRAD_slice_0@GRAD反向梯度命名，当前逻辑会错误解析为z，正确命名应为z@GRAD_slice_0其它参考：