Skip to content

[torchlib] Register aten.linear and use matmul to simplify graph #2021

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
justinchuby merged 4 commits into from
Jan 18, 2025
Merged

[torchlib] Register aten.linear and use matmul to simplify graph #2021

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
de19f32
[torchlib] Register linear and use matmul to simplify graph
justinchuby Jan 18, 2025
61a474e
gemm
justinchuby Jan 18, 2025
77e2a2f
Merge branch 'main' into justinchu/better-linear
justinchuby Jan 18, 2025
32db4bc
test
justinchuby Jan 18, 2025
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
23 changes: 7 additions & 16 deletions onnxscript/function_libs/torch_lib/ops/nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -825,26 +825,17 @@ def aten_leaky_relu_backward(
raise NotImplementedError()


# NOTE: Do not register - We rely on PyTorch decomposition to aten_addmm (Gemm)
def aten_linear(input: TFloat, weight: TFloat) -> TFloat:
@torch_op("aten::linear", trace_only=True)
def aten_linear(input: TFloat, weight: TFloat, bias: TFloat | None = None) -> TFloat:
"""linear(Tensor input, Tensor weight, Tensor? bias=None) -> Tensor"""

# NOTE: The symbolic function in torch.onnx also uses Gemm in certain cases
# Optimizers may consider this path and replace it with Gemm
# We do not use Gemm here because input can have batch dimensions, which Gemm does not support
weight_transposed = op.Transpose(weight, perm=[1, 0])
return op.MatMul(input, weight_transposed)


# NOTE: Do not register - We rely on PyTorch decomposition to aten_addmm (Gemm)
def aten_linear_bias(input: TFloat, weight: TFloat, bias: TFloat) -> TFloat:
"""linear(Tensor input, Tensor weight, Tensor? bias=None) -> Tensor"""

# NOTE: The symbolic function in torch.onnx also uses Gemm in certain cases
# Optimizers may consider this path and replace it with Gemm
# We do not use Gemm here because input can have batch dimensions, which Gemm does not support
if len(input.shape) == 2:
# Use Gemm for the rank 2 input
return op.Gemm(input, weight, bias, transB=True)
weight_transposed = op.Transpose(weight, perm=[1, 0])
mul = op.MatMul(input, weight_transposed)
if bias is None:
return mul
return op.Add(mul, bias)


Expand Down
6 changes: 3 additions & 3 deletions tests/function_libs/torch_lib/ops_test_data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1855,6 +1855,9 @@ def _where_input_wrangler(
tolerance={torch.float16: (8e-2, 1e-4)},
),
TorchLibOpInfo("nn.functional.glu", nn_ops.aten_glu),
TorchLibOpInfo(
"nn.functional.linear", nn_ops.aten_linear, tolerance={torch.float16: (1e-2, 1e-3)}
),
TorchLibOpInfo(
"nn.functional.unfold",
nn_ops.aten_im2col,
Expand Down Expand Up @@ -2176,9 +2179,6 @@ def _where_input_wrangler(
ops_test_common.duplicate_opinfo(OPS_DB, "mean", ("mean_dim",))
ops_test_common.duplicate_opinfo(OPS_DB, "min", ("min_dim",))
ops_test_common.duplicate_opinfo(OPS_DB, "minimum", ("minimum_bool",))
ops_test_common.duplicate_opinfo(
OPS_DB, "nn.functional.linear", ("nn.functional.linear_bias",)
)
ops_test_common.duplicate_opinfo(
OPS_DB,
"nn.functional.pad",
Expand Down
Loading