Skip to content

[mpmd] Do not rename axes in GenerateSdyMeshesFromTopologyPass. #795

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
copybara-service merged 1 commit into from
Oct 2, 2025
Merged

[mpmd] Do not rename axes in GenerateSdyMeshesFromTopologyPass. #795

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
18 changes: 6 additions & 12 deletions shardy/dialect/mpmd/transforms/import/generate_sdy_meshes_from_topology_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,8 +111,7 @@ class GenerateSdyMeshesFromTopologyPass
sharding.getUnreducedAxes());
}
StringRef mesh_name;
auto rename_axes = [&mesh_name](ArrayRef<sdy::AxisRefAttr> axes) {
SmallVector<sdy::AxisRefAttr> new_axes;
auto check_axes = [&mesh_name](ArrayRef<sdy::AxisRefAttr> axes) {
for (sdy::AxisRefAttr axis : axes) {
auto [prefix, axis_name] = axis.getName().split(kMeshAxisSeparator);
SDY_CHECK(!axis_name.empty())
Expand All @@ -123,23 +122,18 @@ class GenerateSdyMeshesFromTopologyPass
<< prefix.str();
}
mesh_name = prefix;
new_axes.push_back(sdy::AxisRefAttr::get(axis.getContext(), axis_name,
axis.getSubAxisInfo()));
}
return new_axes;
};
SmallVector<sdy::DimensionShardingAttr> dim_shardings;
for (auto dim_sharding : sharding.getDimShardings()) {
dim_shardings.push_back(sdy::DimensionShardingAttr::get(
module_op.getContext(), rename_axes(dim_sharding.getAxes()),
dim_sharding.getIsClosed(), dim_sharding.getPriority()));
check_axes(dim_sharding.getAxes());
}
check_axes(sharding.getReplicatedAxes());
check_axes(sharding.getUnreducedAxes());
SDY_CHECK(!llvm::is_contained(old_meshes, mesh_name))
<< "Invalid mesh name: " << mesh_name.str();
return sdy::TensorShardingAttr::get(
sharding.getContext(), mesh_name, dim_shardings,
rename_axes(sharding.getReplicatedAxes()),
rename_axes(sharding.getUnreducedAxes()));
sharding.getContext(), mesh_name, sharding.getDimShardings(),
sharding.getReplicatedAxes(), sharding.getUnreducedAxes());
});

for (StringRef mesh_name : old_meshes) {
Expand Down
60 changes: 30 additions & 30 deletions shardy/dialect/mpmd/transforms/import/test/generate_sdy_meshes_from_topology.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,8 @@

// CHECK-LABEL: module @multiple_input_meshes
module @multiple_input_meshes {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2, "y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["z"=8]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2, "tpu_y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["cpu_z"=8]>
// CHECK-DAG: sdy.mesh @empty_mesh = <[]>
// CHECK-DAG: sdy.mesh @maximal_mesh = <[], device_ids=[0]>
// CHECK-NOT: sdy.mesh @mesh
Expand All @@ -13,16 +13,16 @@ module @multiple_input_meshes {
sdy.mesh @empty_mesh = <[]>
sdy.mesh @maximal_mesh = <[], device_ids=[0]>

// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x", "y"}]>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"z":(1)2}]>}
// CHECK: %arg2: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x", "y"}]>}
// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x", "tpu_y"}]>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"cpu_z":(1)2}]>}
// CHECK: %arg2: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x", "tpu_y"}]>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh_0, [{"cpu_z":(1)2}]>},
%arg2: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg2) (%arg3: tensor<16xf32>, %arg4: tensor<16xf32>) {
%2 = stablehlo.add %arg4, %arg3 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand All @@ -39,20 +39,20 @@ module @multiple_input_meshes {

// CHECK-LABEL: module @empty_mesh
module @empty_mesh {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2]>
// CHECK-DAG: sdy.mesh @empty_mesh = <[]>
// CHECK-NOT: sdy.mesh @mesh
sdy.mesh @mesh = <["tpu_x"=2]>
sdy.mesh @empty_mesh = <[]>

// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x"}]>}
// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x"}]>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@empty_mesh, [{}]>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x"}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@empty_mesh, [{}]>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg1) (%arg2: tensor<16xf32>, %arg3: tensor<16xf32>) {
%2 = stablehlo.add %arg3, %arg2 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand All @@ -65,20 +65,20 @@ module @empty_mesh {

// CHECK-LABEL: module @maximal_mesh
module @maximal_mesh {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2]>
// CHECK-DAG: sdy.mesh @maximal_mesh = <[], device_ids=[0]>
// CHECK-NOT: sdy.mesh @mesh
sdy.mesh @mesh = <["tpu_x"=2]>
sdy.mesh @maximal_mesh = <[], device_ids=[0]>

// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x"}]>}
// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x"}]>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@maximal_mesh, []>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x"}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@maximal_mesh, []>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg1) (%arg2: tensor<16xf32>, %arg3: tensor<16xf32>) {
%2 = stablehlo.add %arg3, %arg2 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand All @@ -91,7 +91,7 @@ module @maximal_mesh {

// CHECK-LABEL: module @fully_replicated_tensor
module @fully_replicated_tensor {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2]>
// CHECK-DAG: sdy.mesh @empty_mesh = <[]>
// CHECK-NOT: sdy.mesh @mesh
sdy.mesh @mesh = <["tpu_x"=2]>
Expand All @@ -102,8 +102,8 @@ module @fully_replicated_tensor {
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{?}]>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2]>>, <"empty_mesh" : <[]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>, <"empty_mesh" : <[]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg1) (%arg2: tensor<16xf32>, %arg3: tensor<16xf32>) {
%2 = stablehlo.add %arg3, %arg2 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand All @@ -116,21 +116,21 @@ module @fully_replicated_tensor {

// CHECK-LABEL: module @replicated_axes
module @replicated_axes {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2, "y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["z"=8]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2, "tpu_y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["cpu_z"=8]>
// CHECK-NOT: sdy.mesh @mesh
// CHECK-NOT: sdy.mesh @mesh_0
sdy.mesh @mesh = <["tpu_x"=8, "tpu_y"=8]>
sdy.mesh @mesh_0 = <["cpu_z"=8]>

// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x"}], replicated={"y"}>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"z":(1)2}]>}
// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x"}], replicated={"tpu_y"}>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"cpu_z":(1)2}]>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x"}], replicated={"tpu_y"}>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh_0, [{"cpu_z":(1)2}]>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg0) (%arg3: tensor<16xf32>, %arg4: tensor<16xf32>) {
%2 = stablehlo.add %arg4, %arg3 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand All @@ -147,21 +147,21 @@ module @replicated_axes {

// CHECK-LABEL: module @unreduced_axes
module @unreduced_axes {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2, "y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["z"=8]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2, "tpu_y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["cpu_z"=8]>
// CHECK-NOT: sdy.mesh @mesh
// CHECK-NOT: sdy.mesh @mesh_0
sdy.mesh @mesh = <["tpu_x"=8, "tpu_y"=8]>
sdy.mesh @mesh_0 = <["cpu_z"=8]>

// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"x"}], unreduced={"y"}>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"z":(1)2}]>}
// CHECK: %arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x"}], unreduced={"tpu_y"}>}
// CHECK: %arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@cpu, [{"cpu_z":(1)2}]>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x"}], unreduced={"tpu_y"}>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh_0, [{"cpu_z":(1)2}]>})
-> (tensor<16xf32>) attributes {
// CHECK: topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>} {
// CHECK: topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>} {
%0 = mpmd.named_computation<"stage1"> (%arg0, %arg0) (%arg3: tensor<16xf32>, %arg4: tensor<16xf32>) {
%2 = stablehlo.add %arg4, %arg3 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand Down
12 changes: 6 additions & 6 deletions shardy/dialect/mpmd/transforms/import/test/import_pipeline_with_heterogeneous_meshes.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
// RUN: mpmd_opt %s -mpmd-import-pipeline='name-to-mesh-assignment=f1@tpu,f2@cpu enable-heterogeneous-meshes' -split-input-file 2>&1 | FileCheck %s

module @multiple_input_meshes {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2, "y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["z"=8]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2, "tpu_y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["cpu_z"=8]>
// CHECK-DAG: sdy.mesh @empty_mesh = <[]>
// CHECK-DAG: sdy.mesh @maximal_mesh = <[], device_ids=[0]>
// CHECK-NOT: sdy.mesh @mesh
Expand All @@ -12,15 +12,15 @@ module @multiple_input_meshes {
sdy.mesh @empty_mesh = <[]>
sdy.mesh @maximal_mesh = <[], device_ids=[0]>

// CHECK: %arg0: !mpmd.mesh_tensor<"tpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@tpu, [{"x", "y"}]>}
// CHECK: %arg1: !mpmd.mesh_tensor<"cpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@cpu, [{"z":(1)2}]>}
// CHECK: %arg2: !mpmd.mesh_tensor<"tpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@tpu, [{"x", "y"}]>}
// CHECK: %arg0: !mpmd.mesh_tensor<"tpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x", "tpu_y"}]>}
// CHECK: %arg1: !mpmd.mesh_tensor<"cpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@cpu, [{"cpu_z":(1)2}]>}
// CHECK: %arg2: !mpmd.mesh_tensor<"tpu", tensor<16xf32>> {sdy.sharding = #sdy.sharding<@tpu, [{"tpu_x", "tpu_y"}]>}
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh_0, [{"cpu_z":(1)2}]>},
%arg2: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>})
-> (tensor<16xf32>) attributes {
topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>} {
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>} {
%0 = mpmd.named_computation<"f1"> (%arg0, %arg2) (%arg3: tensor<16xf32>, %arg4: tensor<16xf32>) {
%2 = stablehlo.add %arg4, %arg3 : tensor<16xf32>
mpmd.return %2 : tensor<16xf32>
Expand Down
12 changes: 6 additions & 6 deletions shardy/dialect/mpmd/transforms/test/e2e_pipeline.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
// RUN: mpmd_opt %s -mpmd-import-pipeline='name-to-mesh-assignment=f1@tpu,f2@cpu enable-heterogeneous-meshes' -mpmd-optimize-pipeline -mpmd-sharding-propagation-pipeline -mpmd-export-pipeline 2>&1 | FileCheck %s

module @multiple_input_meshes {
// CHECK-DAG: sdy.mesh @tpu = <["x"=2, "y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["z"=8]>
// CHECK-DAG: sdy.mesh @tpu = <["tpu_x"=2, "tpu_y"=4]>
// CHECK-DAG: sdy.mesh @cpu = <["cpu_z"=8]>
// CHECK-DAG: sdy.mesh @empty_mesh = <[]>
// CHECK-DAG: sdy.mesh @maximal_mesh = <[], device_ids=[0]>
// CHECK-NOT: sdy.mesh @mesh
Expand All @@ -12,15 +12,15 @@ module @multiple_input_meshes {
sdy.mesh @empty_mesh = <[]>
sdy.mesh @maximal_mesh = <[], device_ids=[0]>

// CHECK: %arg0: !mpmd.mesh_tensor<"tpu", tensor<16xf32>, sharding=<@tpu, [{"x", "y"}]>>
// CHECK: %arg1: !mpmd.mesh_tensor<"cpu", tensor<16xf32>, sharding=<@cpu, [{"z"}]>>
// CHECK: %arg2: !mpmd.mesh_tensor<"tpu", tensor<16xf32>, sharding=<@tpu, [{"x", "y"}]>>
// CHECK: %arg0: !mpmd.mesh_tensor<"tpu", tensor<16xf32>, sharding=<@tpu, [{"tpu_x", "tpu_y"}]>>
// CHECK: %arg1: !mpmd.mesh_tensor<"cpu", tensor<16xf32>, sharding=<@cpu, [{"cpu_z"}]>>
// CHECK: %arg2: !mpmd.mesh_tensor<"tpu", tensor<16xf32>, sharding=<@tpu, [{"tpu_x", "tpu_y"}]>>
func.func @main(
%arg0: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>},
%arg1: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh_0, [{"cpu_z"}]>},
%arg2: tensor<16xf32> {sdy.sharding = #sdy.sharding<@mesh, [{"tpu_x", "tpu_y"}]>})
-> (tensor<16xf32>) attributes {
topology = #mpmd.topology<<"tpu" : <["x"=2, "y"=4]>>, <"cpu" : <["z"=8]>>>} {
topology = #mpmd.topology<<"tpu" : <["tpu_x"=2, "tpu_y"=4]>>, <"cpu" : <["cpu_z"=8]>>>} {
// CHECK: %[[FRAGMENT_CALL1:.*]] = mpmd.fragment_call<mesh="tpu", origin=["f1"]> @p0_f1_fwd.multiple_input_meshes(%arg0, %arg2)
%0 = mpmd.named_computation<"f1"> (%arg0, %arg2) (%arg3: tensor<16xf32>, %arg4: tensor<16xf32>) {
%2 = stablehlo.add %arg4, %arg3 : tensor<16xf32>
Expand Down
Loading