AMDMIGraphX

AMDMIGraphX copied to clipboard

disable const_folding for unpack_int4

lakhinderwalia

Add ONNX parsing for SkipSimplifiedLayerNormalization

3

turneram

richagadgil

[Backport] fix bug in find_concat_op

1

… (#3242) Bugfix needed in torch_migraphx

shivadbhavsar

Introduce Op Builder API

3

- Implement a separate API that performs the graph construction for specific operators instead of it being done directly in the parser - Resolves https://github.com/migraphx-benchmark/AMDMIGraphX/issues/190

mirza-halilcevic

[Issue]: Investigate and Fix GPU error with int8 reduced layer models

2

### Problem Description Seeing GPU fault when running the onnxruntime-inference-examples script using reduced layer bert models during benchmarking. It appears quantization/calibration steps work and the issue arises during inference. ```...

TedThemistokleous

more batch sizes for SD2.1

3

richagadgil

lakhinderwalia

For the interger quantization, quantizelinear operation is ` nearbyint(x / scale) + zeropoint`. FP8 is floating point operation already, Is nearbyint necessary in that case ?

umangyadav

At the time of writing this issue, ONNX has FP8 support for the following ops. ONNX will keep adding FP8 support for more operators. - castlike - cast - constant...

umangyadav