Align NPU to CPU

[KodiaqQ]

Changes

• Aligned NPU configuration to CPU for 8bit quantization;

Reason for changes

• NPU supports CPU configuration;
• CPU configuration as the base;
• Removed unused configurations (ConvolutionBackpropData, GroupConvolutionBackpropData);

Related tickets

• 132512

Tests

• TBD

[codecov[bot]]

Codecov Report

Attention: Patch coverage is 8.33333% with 11 lines in your changes are missing coverage. Please review.

Project coverage is 29.94%. Comparing base (ac15a65) to head (e5cfe06). Report is 1 commits behind head on develop.

Additional details and impacted files

@@             Coverage Diff              @@
##           develop    #2560       +/-   ##
============================================
- Coverage    62.07%   29.94%   -32.13%     
============================================
  Files          494      494               
  Lines        45775    45791       +16     
============================================
- Hits         28413    13714    -14699     
- Misses       17362    32077    +14715     

Files	Coverage Δ
nncf/common/utils/dot_file_rw.py	58.33% <0.00%> (-31.95%)	:arrow_down:
nncf/quantization/algorithms/min_max/algorithm.py	20.17% <9.09%> (-76.10%)	:arrow_down:

... and 262 files with indirect coverage changes

Flag	Coverage Δ
COMMON	?
ONNX	?
OPENVINO	?
TENSORFLOW	29.94% <8.33%> (-0.01%)	:arrow_down:

Flags with carried forward coverage won't be shown. Click here to find out more.

Components	Coverage Δ
common	76.35% <0.00%> (-12.08%)	:arrow_down:
torch	0.01% <ø> (-32.85%)	:arrow_down:
tensorflow	93.74% <ø> (ø)
onnx	0.00% <ø> (-93.07%)	:arrow_down:
openvino	0.00% <ø> (-94.23%)	:arrow_down:
ptq	15.24% <9.09%> (-65.68%)	:arrow_down:

[KodiaqQ]

Linters fails due to https://github.com/openvinotoolkit/nncf/pull/2561

[ljaljushkin]

Should NPU config also contain this?

            "attributes": {
                "scales": "unified"
            },

[ljaljushkin]

Also noticed this difference

[KodiaqQ]

Should NPU config also contain this?

            "attributes": {
                "scales": "unified"
            },

Maybe @alexsu52 and @AlexKoff88 can answer this question.

[KodiaqQ]

Also noticed this difference

What should we do with this? Are these parameters used for NPU?

[alexsu52]

Should NPU config also contain this?

            "attributes": {
                "scales": "unified"
            },

Maybe @alexsu52 and @AlexKoff88 can answer this question.

Good question. Scale unification should be applied only for operation from CPU hardware config to align behavior between CPU and NPU in INT8 quantization. What affects other precision, we should not change their behavior.

[KodiaqQ]

Do you have results of the conformance test for NPU config?

Please think about what test needs to be added to check the equality of the cpu, gpu and npu configs for int8.

We have no tests for NPU in the conformance as well. I haven't run these tests.

[AlexKoff88]

Just for the record, the main motivation for keeping the config is QAT for NPU which has some custom features such as W4A4 support.

[KodiaqQ]

@alexsu52, please, review