[Semi-Auto] Add reduction infer_backward rule #56504
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
Sorry to inform you that bdda6f2's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
pkuzyc
force-pushed
the
reduction_backward
branch
from
bdda6f2
to
a2183fe
Compare
| const paddle::framework::AttributeMap& attrs) override; | ||
|
|
||
| private: | ||
| std::string GetOutputNotation(int64_t input_ndim, |
There was a problem hiding this comment.
the class for rule will be remove and fw & bw will be independent functions.
so place this utils another place and as independent function
|
|
||
| self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [-1, -1]) | ||
| self.assertEqual(infered_output_dist_attrs[0].dims_mapping, [-1]) | ||
|
|
There was a problem hiding this comment.
check the partial is false
| self.assertEqual(infered_input_dist_attrs[0].dims_mapping, [-1, -1]) | ||
| self.assertEqual(infered_output_dist_attrs[0].dims_mapping, [-1]) | ||
|
|
||
| # reduce on dim 0, keep_dim = true |
There was a problem hiding this comment.
should add a kind new case that output is partial on curtain dims, and how it influence input :
stage 1: input is not effect
stage2: input should be sharded on that dims
PR types
Function optimization
PR changes
Others
Description
Pcard-70448
Add infer backward SPMD rule for reduction. It infers inputs' dims mappings as following:
For non-reduced axes, set their dims mappings equal to the outputs'. For the reduced axes, set them to -1.