Skip to content

Use shardy const for sharded constants #825

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
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Use shardy const for sharded constants #825

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
91d9b75
Use shardy const for sharded constants
wsmoses Apr 15, 2025
c5eb57b
fix
wsmoses Apr 15, 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
6 changes: 3 additions & 3 deletions src/enzyme_ad/jax/Passes/OptimizeCommunication.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2265,7 +2265,7 @@ struct DUSToPadComm : public OpRewritePattern<stablehlo::DynamicUpdateSliceOp> {
auto updateType = update.getType().cast<RankedTensorType>();
auto zeroAttr =
DenseElementsAttr::get(updateType, rewriter.getZeroAttr(elementType));
auto zeroUpdateOp = rewriter.create<stablehlo::ConstantOp>(
auto zeroUpdateOp = rewriter.create<sdy::ConstantOp>(
dus.getLoc(), updateType, zeroAttr);
sdy::setSharding(zeroUpdateOp, sharding);

Expand All @@ -2291,7 +2291,7 @@ struct DUSToPadComm : public OpRewritePattern<stablehlo::DynamicUpdateSliceOp> {
} else if (updatePad) {
resultV = updatePad;
} else {
auto cst = rewriter.create<stablehlo::ConstantOp>(
auto cst = rewriter.create<sdy::ConstantOp>(
dus.getLoc(), dus.getType(),
rewriter.getZeroAttr(dus.getType()).cast<ElementsAttr>());
sdy::setSharding(cst, sharding);
Expand Down Expand Up @@ -3353,7 +3353,7 @@ struct ConcatToPadCommOptimize
}

if (addOperands.size() == 0) {
auto cst = rewriter.create<stablehlo::ConstantOp>(
auto cst = rewriter.create<sdy::ConstantOp>(
concat.getLoc(), concat.getType(),
rewriter.getZeroAttr(concat.getType()).cast<ElementsAttr>());
sdy::setSharding(cst, concatSharding);
Expand Down
36 changes: 18 additions & 18 deletions test/lit_tests/communication/dus.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,7 +13,7 @@ func.func @constantUpdate1D(%arg21: tensor<20x24x96xf64> {sdy.sharding = #sdy.sh

// PAD: func.func @constantUpdate1D(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
// PAD-NEXT: %[[cst1:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %[[cst0:.+]] = stablehlo.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<20x24x80xf64>
// PAD-NEXT: %[[cst0:.+]] = sdy.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<20x24x80xf64>
// PAD-NEXT: %[[p0:.+]] = stablehlo.pad %[[cst0]], %[[cst1]], low = [0, 0, 8], high = [0, 0, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<20x24x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[m0:.+]] = stablehlo.multiply %arg0, %[[p0]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %[[m0]] : tensor<20x24x96xf64>
Expand Down Expand Up @@ -60,7 +60,7 @@ func.func @constantUpdate(%arg21: tensor<20x24x96xf64> {sdy.sharding = #sdy.shar

// PAD: func.func @constantUpdate(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
// PAD-NEXT: %[[cst1:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %[[cst0:.+]] = stablehlo.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<4x1x80xf64>
// PAD-NEXT: %[[cst0:.+]] = sdy.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<4x1x80xf64>
// PAD-NEXT: %[[mask:.+]] = stablehlo.pad %[[cst0]], %[[cst1]], low = [8, 8, 8], high = [8, 15, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<4x1x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[mul:.+]] = stablehlo.multiply %arg0, %[[mask]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %[[mul]] : tensor<20x24x96xf64>
Expand Down Expand Up @@ -108,14 +108,14 @@ func.func @argUpdate1D(%arg21: tensor<20x24x96xf64> {sdy.sharding = #sdy.shardin
}

// PAD: func.func @argUpdate1D(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}, %arg1: tensor<20x24x80xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
// PAD-NEXT: %cst = stablehlo.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<20x24x80xf64>
// PAD-NEXT: %cst_0 = stablehlo.constant dense<0.000000e+00> : tensor<f64>
// PAD-NEXT: %cst_1 = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %0 = stablehlo.pad %arg1, %cst_0, low = [0, 0, 8], high = [0, 0, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<20x24x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %1 = stablehlo.pad %cst, %cst_1, low = [0, 0, 8], high = [0, 0, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<20x24x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %2 = stablehlo.multiply %arg0, %1 {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: %3 = stablehlo.add %2, %0 {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %3 : tensor<20x24x96xf64>
// PAD-NEXT: %[[cst0:.+]] = stablehlo.constant dense<0.000000e+00> : tensor<f64>
// PAD-NEXT: %[[cst1:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %[[pad:.+]] = stablehlo.pad %arg1, %[[cst0]], low = [0, 0, 8], high = [0, 0, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<20x24x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[c0:.+]] = sdy.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<20x24x80xf64>
// PAD-NEXT: %[[mask:.+]] = stablehlo.pad %[[c0]], %[[cst1]], low = [0, 0, 8], high = [0, 0, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<20x24x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[mul:.+]] = stablehlo.multiply %arg0, %[[mask]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: %[[res:.+]] = stablehlo.add %[[mul]], %[[pad]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %[[res]] : tensor<20x24x96xf64>
// PAD-NEXT: }

// CHECK: func.func @argUpdate1D(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}, %arg1: tensor<20x24x80xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
Expand Down Expand Up @@ -158,14 +158,14 @@ func.func @argUpdate(%arg21: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<
}

// PAD: func.func @argUpdate(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}, %arg1: tensor<4x1x80xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
// PAD-NEXT: %cst = stablehlo.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<4x1x80xf64>
// PAD-NEXT: %cst_0 = stablehlo.constant dense<0.000000e+00> : tensor<f64>
// PAD-NEXT: %cst_1 = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %0 = stablehlo.pad %arg1, %cst_0, low = [8, 8, 8], high = [8, 15, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<4x1x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %1 = stablehlo.pad %cst, %cst_1, low = [8, 8, 8], high = [8, 15, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<4x1x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %2 = stablehlo.multiply %arg0, %1 {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: %3 = stablehlo.add %2, %0 {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %3 : tensor<20x24x96xf64>
// PAD-NEXT: %[[cst0:.+]] = stablehlo.constant dense<0.000000e+00> : tensor<f64>
// PAD-NEXT: %[[cst1:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<f64>
// PAD-NEXT: %[[pad:.+]] = stablehlo.pad %arg1, %[[cst0]], low = [8, 8, 8], high = [8, 15, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<4x1x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[c0:.+]] = sdy.constant {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} dense<0.000000e+00> : tensor<4x1x80xf64>
// PAD-NEXT: %[[mask:.+]] = stablehlo.pad %[[c0]], %[[cst1]], low = [8, 8, 8], high = [8, 15, 8], interior = [0, 0, 0] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : (tensor<4x1x80xf64>, tensor<f64>) -> tensor<20x24x96xf64>
// PAD-NEXT: %[[mul:.+]] = stablehlo.multiply %arg0, %[[mask]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: %[[res:.+]] = stablehlo.add %[[mul]], %[[pad]] {sdy.sharding = #sdy.sharding_per_value<[<@mesh, [{"z"}, {"y"}, {"x"}]>]>} : tensor<20x24x96xf64>
// PAD-NEXT: return %[[res]] : tensor<20x24x96xf64>
// PAD-NEXT: }

// CHECK: func.func @argUpdate(%arg0: tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}, %arg1: tensor<4x1x80xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) -> (tensor<20x24x96xf64> {sdy.sharding = #sdy.sharding<@mesh, [{"z"}, {"y"}, {"x"}]>}) {
Expand Down
Loading
Loading