Fix fc_mkldnn format issue #38890
Fix fc_mkldnn format issue #38890
Conversation
|
Thanks for your contribution! |
|
Hi @Aganlengzi could you please review? |
| op->Op()->GetAttrIfExists<int>("in_num_col_dims") == 2) { | ||
| q_desc.SetAttr("output_format", Has("data_layout") | ||
| ? Get<std::string>("data_layout") | ||
| : "NCHW"); |
There was a problem hiding this comment.
AddAttr<int>("in_num_col_dims",
"(int, default 1), The fc op can take tensors with more than "
"two dimensions as its inputs.")
.SetDefault(1)
Hi, does it support "in_num_col_dims") == 1 too, cause in_num_col_dims by default is 1. Sorry I did not review clearly last time.
like this ? Any issue?
if (op->Op()->Type() == "fc" &&
op->Op()->GetAttrIfExists<int>("in_num_col_dims") <= 2) {
There was a problem hiding this comment.
There are no issues for fc and in_num_col_dims == 1. The problem is seen only in that narrow cased mentioned in description. For example test_analyzer_int8_resnet50 , which has fc passes enabled and fc op is quantized has no issues.
|
Sorry to inform you that 2dd308f's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually. |
|
@baoachun Could you please review this PR so that we can ask @Aganlengzi to merge it? |
zuzg
dismissed stale reviews from baoachun, lidanqing-intel, and wozna
via
ae9e22d
zuzg
force-pushed
the
fix-fc-mkldnn-format
branch
from
2dd308f to
ae9e22d
Compare
|
Since Baoachun already reviewed this PR and left LGTM hence I just merged this PR |
PR types
Bug fixes
PR changes
OPs
Describe
Adding fc mldnn passes in ernie int8 model caused format error, so the fix is to change NHWC to NCHW in this narrow case. It is a workaround to error "DNNL FC only supports in_num_col_dims equal to 2 when input format is equal to ncw."