Skip to content

Fix slice pattern in fused_flash_attn_pass #63304

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yuanlehome merged 2 commits into from
Apr 8, 2024
Merged

Fix slice pattern in fused_flash_attn_pass #63304

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18d5017
fix slice pattern in fused_flash_attn_pass
yuanlehome Apr 8, 2024
6c25dec
fix typo
yuanlehome Apr 8, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 14 additions & 3 deletions paddle/fluid/pir/transforms/gpu/fused_flash_attn_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -434,15 +434,26 @@ class TransposeSliceFlashAttnPattern : public paddle::drr::DrrPatternBase {
// Result Pattern.
//
paddle::drr::ResultPattern res = src.ResultPattern();
// [b, head, seq_len, head_dim] -> [b, seq_len, head, head_dim]
const auto &q_transpose = res.Op(
"pd_op.transpose", {{"perm", res.VectorInt32Attr({0, 2, 1, 3})}});
res.Tensor("q_transpose") = q_transpose(res.Tensor("q"));
const auto &k_transpose = res.Op(
"pd_op.transpose", {{"perm", res.VectorInt32Attr({0, 2, 1, 3})}});
res.Tensor("k_transpose") = k_transpose(res.Tensor("k"));
const auto &v_transpose = res.Op(
"pd_op.transpose", {{"perm", res.VectorInt32Attr({0, 2, 1, 3})}});
res.Tensor("v_transpose") = v_transpose(res.Tensor("v"));

const auto &flash_attn = res.Op("pd_op.flash_attn",
{{{"dropout", res.Float32Attr(0.0)},
{"causal", res.BoolAttr(false)},
{"return_softmax", res.BoolAttr(false)},
{"is_test", res.BoolAttr(true)},
{"rng_name", res.StrAttr("")}}});
flash_attn({&res.Tensor("q"),
&res.Tensor("k"),
&res.Tensor("v"),
flash_attn({&res.Tensor("q_transpose"),
&res.Tensor("k_transpose"),
&res.Tensor("v_transpose"),
&res.InputNoneTensor(),
&res.Tensor("mask")},
{&res.Tensor("out"),
Expand Down
70 changes: 36 additions & 34 deletions test/ir/pir/fused_pass/test_fused_flash_attn_pass.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -544,41 +544,43 @@ def test_check_output(self):
)
class TestTransposeSliceFlashAttnPattern(PassTest):
r"""
transpose
transpose
|
-----------+----------
| | |
slice slice slice
| | |
Q K V
| | |
| transpose |
| | |
-- matmul-- |
| |
scale |
| |
mask --- add |
| |
softmax |
| |
------matmul------
|
transpose
|
out

transpose
|
------+------
| | |
slice slice slice
| | |
Q K V mask
| | | |
tranpose tranpose tranpose |
| | | |
-------flash_attn------------
|
-----------+----------
| | |
slice slice slice
| | |
Q K V
| | |
| transpose |
| | |
-- matmul-- |
| |
scale |
| |
mask --- add |
| |
softmax |
| |
------matmul------
|
transpose
|
out

transpose
|
------+------
| | |
slice slice slice
| | |
Q K V mask
| | | |
------flash_attn------
|
out
out
"""

def is_program_valid(self, program=None):
Expand Down