Maxim Vafin comments

Results 51 comments of


                                            Maxim Vafin

[Bug]: Argument shapes are inconsistent

Thank you for reporting this issue. It seems to be caused by fusing transformation. @itikhono please take a look.

[JAX FE] Add support for conv_general_dilated.

Is this PR really needed? I see all the relevant changes in #25292

[JAX FE] Add translation for more operations.

> @mvafin Hi, could you please talk more about the implementation of `copy` translation? I tried to return the input directly but the test cannot be passed. Should I do...

[JAX FE] Add translation for more operations.

build_jenkins

[JAX FE] Add translation for more operations.

build_jenkins

Use AvgPool-14 in PT FE

With merge of MaxPool-14 we had performance regression on CPU before it was implemented in plugin and on GPU. Can we test that this change doesn't degrade performance?

[PT FE] Reformat symmetrical GPTQ to use i4 dtype

Accuracy is not the same for this changes. Need to investigate the reason.

[PT FE] Reformat symmetrical GPTQ to use i4 dtype

Issue with accuracy was cause by dynamic quantization which enabled automatically for `i4` but not `u4`. Also fixed zp for asymmetrical case when they ended up in `fp32` instead of...

Support for INT4 decompression patterns from NNCF

build_jenkins

Fix LSTM conversion for models with rank > 3 inputs from Unsqueeze operations

LSTM spec (https://onnx.ai/onnx/operators/onnx__LSTM.html) describes X, init C and H as 3D tensors. Is this a broader behavior supported by onnxruntime?