[MHLO]add native_dropout and related ops pattern #1211
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Yancey1989
requested review from
tanyokwok,
ZihengJiang,
Vremold and
silvasean
silvasean
reviewed
Aug 15, 2022
| @@ -5313,9 +5313,10 @@ def Torch_AtenOnesLikeOp : Torch_Op<"aten.ones_like", [ | |||
| } | |||
|
|
|||
| def Torch_AtenEmptyMemoryFormatOp : Torch_Op<"aten.empty.memory_format", [ | |||
| NoSideEffect, | |||
There was a problem hiding this comment.
Hi, this file is generated and should not be edited manually. Can you please revert this patch so that it does not block other progress, since other people will be running https://github.com/llvm/torch-mlir/blob/main/build_tools/update_torch_ods.sh to update this file and lose your changes and break these tests?
You will need to update this script to make it smart enough to emit these traits:
In this case, I don't think that aten.empty.memory_format is NoSideEffect because it can raise an error if the parameters are wrong. Can you find a way to resubmit this PR without needing this?
There was a problem hiding this comment.
I updated this attribute just because mhlo is value sematic, but it seems aten.empty.memory_format against it. The decomposed pass lowers some patterns such as aten.empty.memory_format + aten.fill, in mhlo sematic, we can lower the fill op to be a constant tensor, instead of creating an empty tensor and fill it. To eliminate the aten.empty.memory_format after Canonicalizer pass, I added the NoSideEffect attribute.
But I agree with you, aten.empty.memory_format may raise an error, I will revert this PR and find another way to lower the native_dropout op.
There was a problem hiding this comment.
Could we change the decomposition to use aten::new_full ?
There was a problem hiding this comment.
It would also be good for us to add a dialect-level canonicalization pattern that erases ReadOnly ops if they don't have uses. That would cover this and many other useful cases.
Yancey1989
added a commit
that referenced
this pull request
Aug 16, 2022
This reverts commit c935795.
Yancey1989
added a commit
to Yancey1989/torch-mlir
that referenced
this pull request
Aug 17, 2022
This reverts commit c935795.
Yancey1989
added a commit
that referenced
this pull request
Aug 17, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Aug 24, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Aug 31, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Aug 31, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Aug 31, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Aug 31, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Sep 1, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Sep 23, 2022
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Sep 23, 2022
qedawkins
pushed a commit
to nod-ai/torch-mlir
that referenced
this pull request
Oct 3, 2022
Signed-off-by: Whitney Tsang <[email protected]> Co-authored-by: Whitney Tsang <[email protected]> Co-authored-by: Ettore Tiotto <[email protected]> Co-authored-by: Tung D. Le <[email protected]>
qedawkins
pushed a commit
to nod-ai/torch-mlir
that referenced
this pull request
Oct 3, 2022
* Add git commit id to llvm.ident metadata * Addressing review comment: llvm#1211 (comment) Signed-off-by: Whitney Tsang <[email protected]> Co-authored-by: Whitney Tsang <[email protected]> Co-authored-by: Ettore Tiotto <[email protected]>
tanyokwok
pushed a commit
that referenced
this pull request
Oct 31, 2022
wyzero
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Nov 4, 2022
* fix float width * fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * reimplement linear lowering * reimplement 2-D rhs for mutmul * add torchdynamo
tanyokwok
pushed a commit
that referenced
this pull request
Nov 10, 2022
* fix float width * fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * add native_dropout and related ops pattern (#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * reimplement linear lowering * reimplement 2-D rhs for mutmul * add torchdynamo
tanyokwok
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Feb 2, 2023
* fix float width * fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * reimplement linear lowering * reimplement 2-D rhs for mutmul * add torchdynamo
JamesTheZ
pushed a commit
to pai-disc/torch-mlir
that referenced
this pull request
Jul 19, 2023
* Fix float width * Fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * Add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * Reimplement linear lowering * Reimplement 2-D rhs for mutmul * Add torchdynamo * Decompose torch.slice_scatter (llvm#1622) * Fix i64 torch.tensor dtype * Add more mhlo basic converters * Alleviate softmax datatype check (#24) * Fix decompose native_batch_norm (#27) * Support group_norm lowering (#25) * Decompose torch.ones/zeros (#28) * Fix softmax output type * Fix gather * Fix some decompose patterns * Not check assert at runtime (#31) * Fix bool tensor attr conversion bug (#32) * Fix mlirDenseElementsAttrBoolGet
JamesTheZ
added a commit
to pai-disc/torch-mlir
that referenced
this pull request
Jul 19, 2023
* Fix float width * Fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * Add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * Reimplement linear lowering * Reimplement 2-D rhs for mutmul * Add torchdynamo * Decompose torch.slice_scatter (llvm#1622) * Fix i64 torch.tensor dtype * Add more mhlo basic converters * Alleviate softmax datatype check (#24) * Fix decompose native_batch_norm (#27) * Support group_norm lowering (#25) * Decompose torch.ones/zeros (#28) * Fix softmax output type * Fix gather * Fix some decompose patterns * Not check assert at runtime (#31) * Fix bool tensor attr conversion bug (#32) * Fix mlirDenseElementsAttrBoolGet Co-Authored-By: ZHENG, Zhen <[email protected]>
JamesTheZ
added a commit
to pai-disc/torch-mlir
that referenced
this pull request
Jul 25, 2023
* Rewrite mhlo with stablehlo after rebase. * Fix BAZEL building error of multiple definition. * Fix float width * Fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * Add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * Reimplement linear lowering * Reimplement 2-D rhs for mutmul * Add torchdynamo * Decompose torch.slice_scatter (llvm#1622) * Fix i64 torch.tensor dtype * Add more mhlo basic converters * Alleviate softmax datatype check (#24) * Fix decompose native_batch_norm (#27) * Support group_norm lowering (#25) * Decompose torch.ones/zeros (#28) * Fix softmax output type * Fix gather * Fix some decompose patterns * Not check assert at runtime (#31) * Fix bool tensor attr conversion bug (#32) * Fix mlirDenseElementsAttrBoolGet --------- Co-authored-by: ZHENG, Zhen <[email protected]>
JamesTheZ
added a commit
to pai-disc/torch-mlir
that referenced
this pull request
Jul 25, 2023
* Rewrite mhlo with stablehlo after rebase. * Fix BAZEL building error of multiple definition. * Fix float width * Fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * Add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * Reimplement linear lowering * Reimplement 2-D rhs for mutmul * Add torchdynamo * Decompose torch.slice_scatter (llvm#1622) * Fix i64 torch.tensor dtype * Add more mhlo basic converters * Alleviate softmax datatype check (#24) * Fix decompose native_batch_norm (#27) * Support group_norm lowering (#25) * Decompose torch.ones/zeros (#28) * Fix softmax output type * Fix gather * Fix some decompose patterns * Not check assert at runtime (#31) * Fix bool tensor attr conversion bug (#32) * Fix mlirDenseElementsAttrBoolGet
JamesTheZ
added a commit
to pai-disc/torch-mlir
that referenced
this pull request
Jul 27, 2023
* Rewrite mhlo with stablehlo after rebase. * Fix BAZEL building error of multiple definition. * Fix float width * Fix divide_floor & export promoteTypes api (#9) * To comply with the old pytorch versions * Add native_dropout_backward & native_layer_norm_backward decomposition (#15) * Add native_dropout and related ops pattern (llvm#1211) * [MHLO] fix dot general contract * Fix batch_norm, div.Tensor_mode and folder (#21) * Reimplement linear lowering * Reimplement 2-D rhs for mutmul * Add torchdynamo * Decompose torch.slice_scatter (llvm#1622) * Fix i64 torch.tensor dtype * Add more mhlo basic converters * Alleviate softmax datatype check (#24) * Fix decompose native_batch_norm (#27) * Support group_norm lowering (#25) * Decompose torch.ones/zeros (#28) * Fix softmax output type * Fix gather * Fix some decompose patterns * Not check assert at runtime (#31) * Fix bool tensor attr conversion bug (#32) * Fix mlirDenseElementsAttrBoolGet
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
See RFC #999
This PR added
aten::native_dropoutop pattern, and some decomposed op patterns, .e.gaten::size.int,aten::valsem.aten.uniform. In thedropout.mlirFilecheck test, this PR used the Mhlo e2e compilation pass pipeline instead of the singleTorchToMhloConvertpass to test the op conversion works as we expected.Co-authored-by: Bairen Yi [email protected]
Co-authored-by: Jiawei Wu [email protected]
Co-authored-by: Tianyou Guo [email protected]
Co-authored-by: Xu Yan [email protected]
Co-authored-by: Ziheng Jiang [email protected]