[Training]

[anna244]

Describe the issue

Reduce-range does not improve the metric

To reproduce

I'm using the reduce-range feature. Quantization is calculated symmetrically, in QDQ format, for int8.

But there are no changes in the metrics. The graphs are completely the same for the INT8 variant and the reduce -range variant. The processor is used together with the instructions avx2, avx512vl, avx512dq, avx512vl, avx512bw, avx512vl. Vinni no.
Green and light blue line is same. green - without reduce - range, light blue -with reduce-range.
Figure 1- we see float 32 and int8 with reduce-range/ figure 2- float 32 and int8 without reduce.

Maybe I need to specify the true flag somewhere else, except in the configuration task. Or perhaps there is an explanation for this behavior of the model.

blue and green line match. Blue line- with reduce-range, green line without.

Urgency

No response

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

1.16.0

PyTorch Version

2.3.1

Execution Provider

openVino

Execution Provider Library Version

No response

Tasks

Beta Give feedback

No tasks being tracked yet.

[xadupre]

cc @yufenglee