Skip to content

[Relax][PyTorch] Fix output shape of torch.nn.functional.scaled_dot_product_attention #17379

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
yongwww merged 3 commits into from
Sep 20, 2024
Merged

[Relax][PyTorch] Fix output shape of torch.nn.functional.scaled_dot_product_attention #17379

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6649037
fix the testcase
mshr-h Sep 16, 2024
f6f6c1a
transpose the output
mshr-h Sep 16, 2024
a2b29c0
fix msc testcase
mshr-h Sep 17, 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
12 changes: 8 additions & 4 deletions python/tvm/contrib/msc/core/transform/pattern.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,15 +330,17 @@ def make_relax_attention_pattern() -> (
q_trans = relax_pattern.is_op("relax.permute_dims")(weight_q)
k_trans = relax_pattern.is_op("relax.permute_dims")(weight_k)
v_trans = relax_pattern.is_op("relax.permute_dims")(weight_v)
out = relax_pattern.is_op("relax.nn.attention")(q_trans, k_trans, v_trans)
attention = relax_pattern.is_op("relax.nn.attention")(q_trans, k_trans, v_trans)
out = relax_pattern.is_op("relax.permute_dims")(attention)
annotations = {
"weight_q": weight_q,
"weight_k": weight_k,
"weight_v": weight_v,
"q_trans": q_trans,
"k_trans": k_trans,
"v_trans": v_trans,
"attention": out,
"attention": attention,
"out": out,
}
return out, annotations

Expand Down Expand Up @@ -378,7 +380,8 @@ def make_relax_mask_attention_pattern() -> (
q_trans = relax_pattern.is_op("relax.permute_dims")(weight_q)
k_trans = relax_pattern.is_op("relax.permute_dims")(weight_k)
v_trans = relax_pattern.is_op("relax.permute_dims")(weight_v)
out = relax_pattern.is_op("relax.nn.attention_bias")(q_trans, k_trans, v_trans, mask)
attention = relax_pattern.is_op("relax.nn.attention_bias")(q_trans, k_trans, v_trans, mask)
out = relax_pattern.is_op("relax.permute_dims")(attention)
annotations = {
"weight_q": weight_q,
"weight_k": weight_k,
Expand All @@ -387,7 +390,8 @@ def make_relax_mask_attention_pattern() -> (
"q_trans": q_trans,
"k_trans": k_trans,
"v_trans": v_trans,
"attention": out,
"attention": attention,
"out": out,
}
return out, annotations

Expand Down
4 changes: 3 additions & 1 deletion python/tvm/relax/frontend/torch/fx_translator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1015,7 +1015,9 @@ def _scaled_dot_product_attention(self, node: fx.Node) -> relax.Var:
assert "float" in attn_mask.struct_info.dtype, msg

return self.block_builder.emit(
relax.op.nn.attention(query, key, value, bias=attn_mask, causal_mask=causal_mask)
transpose_S_H(
relax.op.nn.attention(query, key, value, bias=attn_mask, causal_mask=causal_mask)
)
)

def _unbind(self, node: fx.Node) -> relax.Var:
Expand Down
1 change: 1 addition & 0 deletions src/contrib/msc/framework/tvm/relax_opcode.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,7 @@ class RelaxAttentionCodeGen : public RelaxOpCode {
.op_list_arg<int>(axes_key, "axes");
}
stack_.op_call().op_inputs_arg(false).op_arg<float>("scale").op_str_arg("causal_mask");
stack_.op_call("relax.op.permute_dims").op_output_arg().op_list_arg<int>("axes_3", "axes");
}
};

Expand Down
9 changes: 2 additions & 7 deletions tests/python/contrib/test_msc/test_graph_build.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2362,12 +2362,7 @@ def forward(self, q_data, k_data, v_data):
{"name": "inp_2", "shape": [1, 8, seq, 64], "dtype": "float32", "layout": "ACBD"},
],
"outputs": [
{
"name": "attention",
"shape": [1, seq, 8, 64],
"dtype": "float32",
"layout": "ABCD",
}
{"name": "attention", "shape": [1, 8, seq, 64], "dtype": "float32", "layout": "ABCD"}
],
"nodes": {"total": 4, "input": 3, "msc.attention": 1},
}
Expand Down Expand Up @@ -2396,7 +2391,7 @@ def forward(self, q_data, k_data, v_data, mask):
"outputs": [
{
"name": "attention_bias",
"shape": [1, seq, 8, 64],
"shape": [1, 8, seq, 64],
"dtype": "float32",
"layout": "ABCD",
}
Expand Down
21 changes: 15 additions & 6 deletions tests/python/relax/test_frontend_from_fx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3825,7 +3825,7 @@ def main(
inp_0: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_1: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_2: R.Tensor((32, 8, 128, 64), dtype="float32"),
) -> R.Tensor((32, 128, 8, 64), dtype="float32"):
) -> R.Tensor((32, 8, 128, 64), dtype="float32"):
with R.dataflow():
lv: R.Tensor((32, 128, 8, 64), dtype="float32") = R.permute_dims(
inp_0, axes=[0, 2, 1, 3]
Expand All @@ -3839,7 +3839,10 @@ def main(
lv3: R.Tensor((32, 128, 8, 64), dtype="float32") = R.nn.attention(
lv, lv1, lv2, scale=None
)
gv: R.Tensor((32, 128, 8, 64), dtype="float32") = lv3
lv4: R.Tensor((32, 8, 128, 64), dtype="float32") = R.permute_dims(
lv3, axes=[0, 2, 1, 3]
)
gv: R.Tensor((32, 8, 128, 64), dtype="float32") = lv4
R.output(gv)
return gv

Expand All @@ -3851,7 +3854,7 @@ def main(
inp_1: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_2: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_3: R.Tensor((32, 8, 128, 128), dtype="float32"),
) -> R.Tensor((32, 128, 8, 64), dtype="float32"):
) -> R.Tensor((32, 8, 128, 64), dtype="float32"):
with R.dataflow():
lv: R.Tensor((32, 128, 8, 64), dtype="float32") = R.permute_dims(
inp_0, axes=[0, 2, 1, 3]
Expand All @@ -3865,7 +3868,10 @@ def main(
lv3: R.Tensor((32, 128, 8, 64), dtype="float32") = R.nn.attention(
lv, lv1, lv2, inp_3, scale=None
)
gv: R.Tensor((32, 128, 8, 64), dtype="float32") = lv3
lv4: R.Tensor((32, 8, 128, 64), dtype="float32") = R.permute_dims(
lv3, axes=[0, 2, 1, 3]
)
gv: R.Tensor((32, 8, 128, 64), dtype="float32") = lv4
R.output(gv)
return gv

Expand All @@ -3876,7 +3882,7 @@ def main(
inp_0: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_1: R.Tensor((32, 8, 128, 64), dtype="float32"),
inp_2: R.Tensor((32, 8, 128, 64), dtype="float32"),
) -> R.Tensor((32, 128, 8, 64), dtype="float32"):
) -> R.Tensor((32, 8, 128, 64), dtype="float32"):
with R.dataflow():
lv: R.Tensor((32, 128, 8, 64), dtype="float32") = R.permute_dims(
inp_0, axes=[0, 2, 1, 3]
Expand All @@ -3890,7 +3896,10 @@ def main(
lv3: R.Tensor((32, 128, 8, 64), dtype="float32") = R.nn.attention(
lv, lv1, lv2, scale=None, causal_mask="TopLeft"
)
gv: R.Tensor((32, 128, 8, 64), dtype="float32") = lv3
lv4: R.Tensor((32, 8, 128, 64), dtype="float32") = R.permute_dims(
lv3, axes=[0, 2, 1, 3]
)
gv: R.Tensor((32, 8, 128, 64), dtype="float32") = lv4
R.output(gv)
return gv

Expand Down