未验证 提交 cbda49e6 编写于 作者: J joanna.wozna.intel 提交者: GitHub

Correct order of passes (#45936)

上级 1711407d
......@@ -448,13 +448,14 @@ class Quant2Int8MkldnnPass(object):
graph = self._apply_pass(graph, 'fc_act_mkldnn_fuse_pass')
graph = self._apply_pass(graph,
'matmul_transpose_reshape_mkldnn_fuse_pass')
graph = self._apply_pass(graph,
'matmul_elementwise_add_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'matmul_activation_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'batch_norm_act_fuse_pass')
graph = self._apply_pass(graph, 'softplus_activation_mkldnn_fuse_pass')
graph = self._apply_pass(graph, 'scale_matmul_fuse_pass')
graph = self._apply_pass(graph,
'reshape_transpose_matmul_mkldnn_fuse_pass')
graph = self._apply_pass(graph,
'matmul_elementwise_add_mkldnn_fuse_pass')
# the following pass should be the last one since it will work on all fused ops.
graph = self._apply_pass(graph, 'runtime_context_cache_pass')
return graph
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册