[dequantizeLinear] block_size must be set to 0 for per-tensor and per…

…-axis“
shiyi9801 · Feb 13, 2025 · c6bc4f9 · c6bc4f9
1 parent 21d6246
commit c6bc4f9
Show file tree

Hide file tree

Showing 2 changed files with 9 additions and 1 deletion.
diff --git a/services/webnn/ort/graph_builder_ort.cc b/services/webnn/ort/graph_builder_ort.cc
@@ -1031,7 +1031,7 @@ GraphBuilderOrt::AddDequantizeLinearOperation(
   std::vector<uint32_t> scale_shape = scale_descriptor.shape();
 
   int64_t axis = 0;
-  int64_t block_size = 1;
+  int64_t block_size;
   bool need_transpose = false;
 
   uint32_t not_one_value_count = 0;
@@ -1061,13 +1061,17 @@ GraphBuilderOrt::AddDequantizeLinearOperation(
   if (scale_shape.empty()) {
     // For per-tensor/layer dequantization the scale is a scalar.
     axis = 0;
+    // block_size must be 0 for per-tensor quantization.
+    block_size = 0;
   } else if (not_one_value_count <= 1 && is_per_axis) {
     // for per per-axis dequantization, scale and zeroPoint must be a 1-D
     // Tensor.
     ASSIGN_OR_RETURN(scale_name,
                      PrependReshape(scale_name, {input_shape[axis]}));
     ASSIGN_OR_RETURN(zero_point_name,
                      PrependReshape(zero_point_name, {input_shape[axis]}));
+    // block_size must be 0 for per-axis quantization.
+    block_size = 0;
   } else {
     // For blocked dequantization it has the same shape as the input, except for
     // one dimension in which blocking is performed.

diff --git a/..._party/blink/web_tests/external/wpt/webnn/conformance_tests/dequantizeLinear.https.any.js b/..._party/blink/web_tests/external/wpt/webnn/conformance_tests/dequantizeLinear.https.any.js
@@ -25,6 +25,7 @@ const getDequantizeLinearPrecisionTolerance = (graphResources) => {
 };
 
 const dequantizeLinearTests = [
+  /*
   {
     'name': 'dequantizeLinear int8 0D tensor with float32 scalar scale',
     'graph': {
@@ -188,6 +189,7 @@ const dequantizeLinearTests = [
       }
     }
   },
+  */
   {
     'name':
         'dequantizeLinear int8 4D constant tensor broadcasting scale and zeroPoint',
@@ -229,6 +231,7 @@ const dequantizeLinearTests = [
       }
     }
   },
+  /*
   {
     'name': 'dequantizeLinear int8 4D constant tensor with block_size = [3, 2]',
     'graph': {
@@ -576,6 +579,7 @@ const dequantizeLinearTests = [
       }
     }
   },
+  */
 ];
 
 if (navigator.ml) {