PawelGlomski-Intel commented on a change in pull request #20753:
URL: https://github.com/apache/incubator-mxnet/pull/20753#discussion_r828355834
##########
File path: src/operator/subgraph/dnnl/dnnl_fc.cc
##########
@@ -662,7 +662,11 @@ static bool SgDNNLFCInferType(const nnvm::NodeAttrs& attrs,
}
return true;
} else {
- return DefaultSubgraphOpType(attrs, in_types, out_types);
+ bool result = DefaultSubgraphOpType(attrs, in_types, out_types);
+ if (full_param.dnnl_param.amp_out_dtype.has_value()) {
Review comment:
> Is it means that we assume float32 type?
`amp_out_dtype` can be any dtype (that was in the fused `amp_cast` node)
that the `ONEDNN_AMP` pass decided that this op can fuse. Currently, only
float32 can be fused.
> Also quantized version could produce float if enable_float_output is set.
Could it be writen as bfloat ?
No, fp32 ops currently cannot return bfloat16 outputs.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
