From 1aa6db4aafc7a6c961b82b5cadee44c45927de6c Mon Sep 17 00:00:00 2001
From: Vladimir Zinoviev <vladimir.zinoviev@intel.com>
Date: Wed, 22 Sep 2021 01:15:36 +0300
Subject: [PATCH] [LPT] Move FQ disabled for non-per-tensor FQs (#7581)

---
 .../src/move_fake_quantize.cpp                   | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/inference-engine/src/low_precision_transformations/src/move_fake_quantize.cpp b/inference-engine/src/low_precision_transformations/src/move_fake_quantize.cpp
index 7192282c2d2..a64d6c66f79 100644
--- a/inference-engine/src/low_precision_transformations/src/move_fake_quantize.cpp
+++ b/inference-engine/src/low_precision_transformations/src/move_fake_quantize.cpp
@@ -39,12 +39,26 @@ MoveFakeQuantize::MoveFakeQuantize(const Params& params) : LayerTransformation(p
         output_low,
         output_high });
 
-    ngraph::graph_rewrite_callback callback = [this](pattern::Matcher& m) {
+    ngraph::graph_rewrite_callback callback = [=](pattern::Matcher& m) {
         auto op = m.get_match_root();
         if (transformation_callback(op)) {
             return false;
         }
 
+        // workaround: only per-tensor quantization is allowed
+        const auto& pattern_map = m.get_pattern_value_map();
+        const auto is_scalar = [&](const std::shared_ptr<ngraph::Node>& wrapped_constant) {
+            return NetworkHelper::isScalarLike(
+                    as_type_ptr<opset1::Constant>(pattern_map.at(wrapped_constant).get_node_shared_ptr()));
+        };
+
+        if (!is_scalar(input_low) ||
+            !is_scalar(input_high) ||
+            !is_scalar(output_low) ||
+            !is_scalar(output_high)) {
+            return false;
+        }
+
         return transform(*context, m);
     };