From cb6c0e213339a59728e9571d3dc62686e2acf879 Mon Sep 17 00:00:00 2001
From: XGZhang <46363693+XGZhang11@users.noreply.github.com>
Date: Thu, 4 Nov 2021 19:48:31 +0800
Subject: [PATCH] Fix a bug of quantization (#36982)

* fix a quantization bug
---
 .../fluid/contrib/slim/quantization/quantization_pass.py     | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
index 90caee6c7a..9b2954b13f 100644
--- a/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
+++ b/python/paddle/fluid/contrib/slim/quantization/quantization_pass.py
@@ -1292,10 +1292,11 @@ class QuantizationFreezePass(object):
             var_type=output_var_node.type(),
             shape=output_var_node.shape(),
             var_dtype=output_var_node.dtype())
+        x_num_col_dims = 1
+        if op_node.name() in ['matmul', 'matmul_v2', 'mul']:
+            x_num_col_dims = len(op_node.outputs[0].shape()) - 1
         if op_node.op().has_attr("x_num_col_dims"):
             x_num_col_dims = op_node.op().attr("x_num_col_dims")
-        else:
-            x_num_col_dims = 1
         dequant_op_node = graph.create_op_node(
             op_type='fake_channel_wise_dequantize_max_abs',
             attrs={
-- 
GitLab