[CINN] 【Infer Symbolic Shape BUAA 】Add flashmask_attention op #68385
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
crazyxiaoxi
changed the title
| @@ -1543,6 +1543,84 @@ bool FlashAttnOpInferSymbolicShape( | |||
| // return true; | |||
| // } | |||
|
|
|||
| bool FlashmaskAttentionOpInferSymbolicShape( | |||
There was a problem hiding this comment.
对比下和 FlashAttnOpInferSymbolicShape 的区别，看看是不是能直接复用
There was a problem hiding this comment.
两个函数在qkv的处理和output的计算是一样的。
但是输入参数所取的operand位置不一样，同时输入的参数也不太相同，不能直接复用。
Comment on lines
1574
to
1582
| PADDLE_ENFORCE_EQ( | ||
| infer_context->IsEqual(startend_row_indices[3], symbol::DimExpr{1}) || | ||
| infer_context->IsEqual(startend_row_indices[3], | ||
| symbol::DimExpr{2}) || | ||
| infer_context->IsEqual(startend_row_indices[3], symbol::DimExpr{4}), | ||
| true, | ||
| common::errors::InvalidArgument( | ||
| "flashmask_attention startend_row_indices " | ||
| "mask_bounds must in [1,2,4]")); |
There was a problem hiding this comment.
这个enforce先删掉吧，目前还不支持编译期在控制流里判等。要写的话需要先判断下startend_row_indices[3]是不是int类型
Comment on lines
+1598
to
+1622
| auto batch_size_expr = q.shape()[0]; | ||
| auto num_heads_expr = q.shape()[2]; | ||
| auto seqlen_q_rounded_expr = round_multiple(q.shape()[1]); | ||
| auto seqlen_k_rounded_expr = round_multiple(k.shape()[1]); | ||
|
|
||
| if (op->result(1)) { | ||
| std::vector<symbol::DimExpr> softmax_shape{batch_size_expr, | ||
| num_heads_expr, | ||
| seqlen_q_rounded_expr, | ||
| seqlen_k_rounded_expr}; | ||
| infer_context->SetShapeOrDataForValue( | ||
| op->result(1), symbol::TensorShapeOrDataDimExprs(softmax_shape)); | ||
| } | ||
| if (op->result(2)) { | ||
| std::vector<symbol::DimExpr> softmax_lse_shape{ | ||
| batch_size_expr, num_heads_expr, seqlen_q_rounded_expr}; | ||
| infer_context->SetShapeOrDataForValue( | ||
| op->result(2), symbol::TensorShapeOrDataDimExprs(softmax_lse_shape)); | ||
| } | ||
| if (op->result(3)) { | ||
| std::vector<symbol::DimExpr> seed_offset_shape{symbol::DimExpr{2}}; | ||
| infer_context->SetShapeOrDataForValue( | ||
| op->result(3), symbol::TensorShapeOrDataDimExprs(out_shape)); | ||
| } | ||
| return true; |
There was a problem hiding this comment.
这部分推导逻辑没在kernel、infermeta、以及FlashAttnOpInferSymbolicShape中找到对应依据
|
Sorry to inform you that 266fe33's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
Hongqing-work
approved these changes
Oct 14, 2024
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.
PR Category
CINN
PR Types
improvements
Description
添加flashmask_attention算子符号推导接口
找到单测
/test/legacy_test/test_flashmask.py
/test/legacy_test/test_flash_attention.py
只有unnitest
Pcard-67164