[Auto Parallel] Add spmd rule No.15 for index_put and index_put_grad ops. #73486
Conversation
| 你的PR提交成功，感谢你对开源项目的贡献! |
| /re-run all-failed |
| // build input data class | ||
| std::vector<int64_t> x_shape = {64, 64, 64}; | ||
| std::vector<int64_t> indice_shape = {32}; | ||
| std::vector<int64_t> value_shape = {32}; |
There was a problem hiding this comment.
in this case, if value_shape=[32, 64], can run normally on single device, but can run normally on multi device? especially dims_mapping of x is {-1, 0, 1}
tips: After using this spmd rule, the local shape will be: x_shape=[64, 64, 32], indice_shape=([32], [32]), value_shape=[32, 64], Can index_put operator will run successfully?
There was a problem hiding this comment.
Thanks, I fix this situation(value_shape greater than 1) in next commit.
| x_ndim, | ||
| indices_size)); | ||
| | ||
| PADDLE_ENFORCE_EQ(value_ndim, |
There was a problem hiding this comment.
According to the definition of the index_put operator, it can support value_ndim greater than 1. This restriction cannot be applied as it can be verified through the following code:
x = paddle.zeros([4, 4, 4]) value = paddle.ones([4, 4]) # can support value_ndim > 1(meet certain conditions) # value = paddle.arange(16).reshape([4, 4]).astype("float32") ix1 = paddle.to_tensor([0,1,2,3]) ix2 = paddle.to_tensor([0,1,2,3]) indices=(ix1,ix2) out = paddle.index_put(x,indices,value) print(out) # can run normallyThere was a problem hiding this comment.
Thanks, I have removed this restriction.
| /re-run all-failed |
| /re-run all-failed |
| } | ||
| } else { // indices_size = x_ndim | ||
| for (int i = 0; i < x_ndim; i++) { | ||
| x_dims_mapping[i] = -1; |
There was a problem hiding this comment.
the branch of else only handles the case of indices_size = x_ndim? can it be merged into the above branch, so we do not need to distinguish between if and else
There was a problem hiding this comment.
Thanks for your suggestion, I removed the if and else in this code block.
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@ ## develop #73486 +/- ## =========================================== Coverage ? 100.00% =========================================== Files ? 1 Lines ? 122 Branches ? 0 =========================================== Hits ? 122 Misses ? 0 Partials ? 0 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
5 participants
PR Category
Auto Parallel
PR Types
New features
Description