54c4e350ab
### Details: `ConvolutionTransformation` propagates multiplies from data and weights paths, and then fuses these 2 scales into one. The bug was In mixed precision scenario (which is used on GPU device) when data dequantization was not produced by LPT but existed in the model. The propagated scales from data and weights had different precisions: scale from data did not change precision (f16), whereas precision of scale from weights was set to `deqPrecision=f32` - this led to element type inconsistency error at the scales fusion stage. In this PR, both scales are converted to `deqPrecision` to align scales element types and avoid precisions incosistency. The same issue existed in `ConvolutionBackpropDataTransformation`: it was fixed too. ### Tickets: - *CVS-130598* |
||
|---|---|---|
| .. | ||
| include | ||
| src | ||
| tests | ||
| CMakeLists.txt | ||