Skip to content

[Relax] Skip 1-D inputs in CombineParallelMatmul #20389

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

Open
Nanmur wants to merge 1 commit into
base: main
Choose a base branch
from
+44 −2
Open

[Relax] Skip 1-D inputs in CombineParallelMatmul #20389

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
8 changes: 6 additions & 2 deletions src/relax/transform/combine_parallel_matmul.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,9 +145,14 @@ ffi::TypedFunction<ffi::Map<Var, Expr>(ffi::Map<DFPattern, Var>, ffi::Map<Var, E
ffi::Map<Var, Expr> replacements;

for (const auto& [rhs_dim, indices] : GroupShapes(rhs_shapes)) {
if (indices.size() == 1 || !batch_dims_compatible(rhs_dim, indices, rhs_shapes)) continue;
if (rhs_dim < 2 || indices.size() == 1 ||
!batch_dims_compatible(rhs_dim, indices, rhs_shapes)) {
continue;
}

auto lhs = matchings[patterns.input];
int lhs_dim = GetTensorType(lhs)->ndim;
if (lhs_dim < 2) continue;

const auto& patterns_to_replace = [&patterns, &branch_info]() {
if (branch_info.activation) return patterns.activation;
Expand Down Expand Up @@ -243,7 +248,6 @@ ffi::TypedFunction<ffi::Map<Var, Expr>(ffi::Map<DFPattern, Var>, ffi::Map<Var, E
sections.push_back(IntImm::Int64(split_index));
}

int lhs_dim = GetTensorType(lhs)->ndim;
int split_axis = std::max<int>(lhs_dim, rhs_dim) - 1;
auto chunks = split(matmul_combined, sections, split_axis);

Expand Down
38 changes: 38 additions & 0 deletions tests/python/relax/test_transform_combine_parallel_matmul.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -733,5 +733,43 @@ def before(
tvm.ir.assert_structural_equal(after, before)


def test_skip_vector_lhs():
@R.function(private=True)
def before(
x: R.Tensor((16,), "float32"),
w0: R.Tensor((16, 32), "float32"),
w1: R.Tensor((16, 32), "float32"),
):
with R.dataflow():
y0 = R.matmul(x, w0)
y1 = R.matmul(x, w1)
out = (y0, y1)
R.output(out)
return out

after = CombineParallelMatmul()(tvm.IRModule.from_expr(before))["main"]

tvm.ir.assert_structural_equal(after, before)


def test_skip_vector_rhs():
@R.function(private=True)
def before(
x: R.Tensor((8, 16), "float32"),
w0: R.Tensor((16,), "float32"),
w1: R.Tensor((16,), "float32"),
):
with R.dataflow():
y0 = R.matmul(x, w0)
y1 = R.matmul(x, w1)
out = (y0, y1)
R.output(out)
return out

after = CombineParallelMatmul()(tvm.IRModule.from_expr(before))["main"]

tvm.ir.assert_structural_equal(after, before)


if __name__ == "__main__":
tvm.testing.main()
Loading