Skip to content

[Unity][FX] Add support for PT2.0 scaled_dot_product_attention #14841

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
vinx13 merged 7 commits into from
May 16, 2023
Merged

[Unity][FX] Add support for PT2.0 scaled_dot_product_attention #14841

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
e22cad9
add converter for PT 2.0 scaled_dot_product_attention
masahi May 12, 2023
918a046
remove requires_gpu in FX test
masahi May 12, 2023
970e532
add test
masahi May 12, 2023
7dc449e
more black
masahi May 12, 2023
d444388
support float mask for attn_mask input
masahi May 14, 2023
bc4b5b4
remove local import
masahi May 15, 2023
19dad73
more clean
masahi May 15, 2023
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
18 changes: 18 additions & 0 deletions 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 @@ -1002,6 +1002,23 @@ def _cross_entropy(self, node: fx.node.Node) -> relax.Expr:
)
)

def _scaled_dot_product_attention(self, node: fx.node.Node) -> relax.Var:
assert len(node.args) <= 4, "Dropout, and causal masking are not supported."
transpose_S_H = lambda tensor: relax.op.permute_dims(tensor, [0, 2, 1, 3])
query = transpose_S_H(self.env[node.args[0]])
key = transpose_S_H(self.env[node.args[1]])
value = transpose_S_H(self.env[node.args[2]])

if len(node.args) == 4:
mask = self.env[node.args[3]]
msg = "Only a float mask is supported for the attn_mask input."
assert "float" in mask.struct_info.dtype, msg
attn = relax.op.nn.attention(query, key, value, bias=mask)
else:
attn = relax.op.nn.attention(query, key, value)

return self.block_builder.emit(attn)

########## Others ##########

def _size(self, node: fx.node.Node) -> relax.Expr:
Expand Down Expand Up @@ -1185,6 +1202,7 @@ def create_convert_map(self):
"neg": self._neg,
"max": self._max,
"cross_entropy": self._cross_entropy,
"scaled_dot_product_attention": self._scaled_dot_product_attention,
}

def from_fx(
Expand Down
Loading