Skip to content

[IE] Mark constant 2-bit weights for the DynDeq -> FakeQuantize bridge in multi-row graphs #390

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
ryugyosoft-max wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

[IE] Mark constant 2-bit weights for the DynDeq -> FakeQuantize bridge in multi-row graphs #390

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
bf06548
Mark constant 2-bit weights for the DynDeq -> FakeQuantize bridge in …
ryugyosoft-max Oct 28, 2026
b1376e6
Update the LIT tests of the constant 2-bit prefill pattern
ryugyosoft-max Oct 28, 2026
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
4 changes: 3 additions & 1 deletion ...ux_compiler/src/dialect/IE/transforms/passes/weights_dequantize_to_dynamic_dequantize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,7 +113,9 @@ mlir::LogicalResult finishDynamicDequantizeRewrite(mlir::PatternRewriter& rewrit
rewriter.create<IE::DynamicDequantizeOp>(appendLoc(loc, "ddq"), rawWeights, scaleValue, zpValue, dstType);

// Mark the op for the DynDeq -> FakeQuantize bridge pass.
if (wdInfo.isKVcachedPattern() || !zpElemType.isInteger(2)) {
// Constant 2-bit weights are marked in the multi-row (prefill) pattern too: left unmarked, they are dequantized by
// the DynamicDequantize SW kernel, which returns zeros for them on NPU4000 / NPU5010 (openvino#39010).
if (wdInfo.isKVcachedPattern() || !zpElemType.isInteger(2) || wdInfo.hasConstWeights()) {
dynamicDequant->setAttr(IE::WEIGHTS_IMPORT_DYN_DEQUANT_ATTR, mlir::UnitAttr::get(rewriter.getContext()));
}

Expand Down
39 changes: 37 additions & 2 deletions tests/lit/NPU/dialect/IE/passes/weights_dequantize_to_dynamic_dequantize.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -697,8 +697,7 @@ func.func @U2WeightsConstPrefillPatternToDynamicDequantize(%act : tensor<1x3x4xf
// CHECK-DAG: [[WT_RAW:%.+]] = const.Declare tensor<4x2x2xui2> = dense_resource<weights_blob>
// CHECK-DAG: [[ZP:%.+]] = const.Declare tensor<4x2x1xui2> = dense_resource<shift_blob>

// CHECK-NOT: vpux.weights_import_dyn_dequant
// CHECK: [[DDQ:%.+]] = IE.DynamicDequantize([[WT_RAW]], [[SCALE]], [[ZP]]) {dstElemType = f16} : tensor<4x2x2xui2>, tensor<4x2x1xf16>, tensor<4x2x1xui2> -> tensor<4x2x2xf16>
// CHECK: [[DDQ:%.+]] = IE.DynamicDequantize([[WT_RAW]], [[SCALE]], [[ZP]]) {dstElemType = f16, vpux.weights_import_dyn_dequant} : tensor<4x2x2xui2>, tensor<4x2x1xf16>, tensor<4x2x1xui2> -> tensor<4x2x2xf16>
// CHECK: [[RESHAPE_1:%.+]] = IE.AffineReshape([[DDQ]])
// CHECK: [[CONVERT:%.+]] = IE.Convert([[RESHAPE_1]]) {dstElemType = f32} : tensor<4x4xf16> -> tensor<4x4xf32>
// CHECK: [[RESHAPE_2:%.+]] = IE.Reshape([[ACT]]) {shape_value = [3, 4]} : tensor<1x3x4xf32> -> tensor<3x4xf32>
Expand All @@ -718,6 +717,42 @@ func.func @U2WeightsConstPrefillPatternToDynamicDequantize(%act : tensor<1x3x4xf

// -----

// CHECK-LABEL: @I2WeightsConstPrefillPatternToDynamicDequantize
// CHECK-SAME: [[ACT:%.+]]: tensor<1x3x4xf32>
func.func @I2WeightsConstPrefillPatternToDynamicDequantize(%act : tensor<1x3x4xf32>) -> tensor<1x3x4xf32> {
%weights = const.Declare tensor<4x2x2xf16> = dense_resource<weights_blob> : tensor<16xsi2>, [#const.ConvertElemType<si8>, #const.Reshape<[4, 2, 2]>, #const.CastElemType<f16>]
%scale = const.Declare tensor<4x2x1xf16> = dense<[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8]> : tensor<8xf16>, [#const.Reshape<[4, 2, 1]>]

%1 = IE.Multiply(%weights, %scale) {auto_broadcast = #IE.auto_broadcast_type<NUMPY>} : tensor<4x2x2xf16>, tensor<4x2x1xf16> -> tensor<4x2x2xf16>
%2 = IE.AffineReshape(%1) {dim_mapping = [[0], [1], [1]], shape_value = [4, 4]} : tensor<4x2x2xf16> -> tensor<4x4xf16>
%3 = IE.Convert(%2) {dstElemType = f32} : tensor<4x4xf16> -> tensor<4x4xf32>
%4 = IE.Reshape(%act) {shape_value = [3, 4]} : tensor<1x3x4xf32> -> tensor<3x4xf32>
%5 = IE.FullyConnected(%4, %3) : tensor<3x4xf32>, tensor<4x4xf32> -> tensor<3x4xf32>
%6 = IE.Reshape(%5) {shape_value = [1, 3, 4]} : tensor<3x4xf32> -> tensor<1x3x4xf32>
return %6 : tensor<1x3x4xf32>

// CHECK-DAG: [[SCALE:%.+]] = const.Declare tensor<4x2x1xf16>
// CHECK-DAG: [[WT_RAW:%.+]] = const.Declare tensor<4x2x2xsi2> = dense_resource<weights_blob>

// CHECK: [[DDQ:%.+]] = IE.DynamicDequantize([[WT_RAW]], [[SCALE]]) {dstElemType = f16, vpux.weights_import_dyn_dequant} : tensor<4x2x2xsi2>, tensor<4x2x1xf16> -> tensor<4x2x2xf16>
// CHECK: [[RESHAPE_1:%.+]] = IE.AffineReshape([[DDQ]])
// CHECK: [[CONVERT:%.+]] = IE.Convert([[RESHAPE_1]]) {dstElemType = f32} : tensor<4x4xf16> -> tensor<4x4xf32>
// CHECK: [[RESHAPE_2:%.+]] = IE.Reshape([[ACT]]) {shape_value = [3, 4]} : tensor<1x3x4xf32> -> tensor<3x4xf32>
// CHECK: [[FC:%.+]] = IE.FullyConnected([[RESHAPE_2]], [[CONVERT]]) : tensor<3x4xf32>, tensor<4x4xf32> -> tensor<3x4xf32>
// CHECK: [[RESHAPE_3:%.+]] = IE.Reshape([[FC]]) {shape_value = [1, 3, 4]} : tensor<3x4xf32> -> tensor<1x3x4xf32>
// CHECK: return [[RESHAPE_3]]
}

{-#
dialect_resources: {
builtin: {
weights_blob: "0x010000001AA12345"
}
}
#-}

// -----

// CHECK-LABEL: @DontLhsConstScaleToDynamicDequantize
// CHECK-SAME: [[INPUT:%.+]]: tensor<1x2xf32>
func.func @DontLhsConstScaleToDynamicDequantize(%arg0: tensor<1x2xf32>) -> tensor<1x6xf32> {
Expand Down
Loading