[PDPD FE] : Refactored PaddlePaddle Quantization

[Pey-crypto]

Fix for #20687

1. Supports both per-tensor and per-channel quantization using the axis attribute.
2. Uses FakeQuantize for quantize_linear and a simple subtract-multiply operation for dequantize_linear.
3. The HALF_AWAY_FROM_ZERO round mode is ignored, as it's not explicitly handled.

[xczhai]

Hi @Pey-crypto
Thanks for your commits. But we can polish this PR for better maintain.

1. please check https://github.com/openvinotoolkit/openvino/blob/master/src/frontends/paddle/src/internal/pass/transform_fakequantize.cpp is needed or not. Theoretically, such pass should be remove since your commit works.
2. Previously, I prepare some unittests for Quantization. Maybe you can import these tests to such PR. It can ensure the motification accuracy. https://github.com/xczhai/openvino/tree/refactor_paddle_slim

BR.

[Pey-crypto]

@xczhai ,
Currently working on the errors.

[xczhai]

@xczhai , Currently working on the errors.

If you have any questions, feel free reach me. I will reply asap.

[Pey-crypto]

@xczhai , Currently working on the errors.

If you have any questions, feel free reach me. I will reply asap.

https://github.com/xczhai/openvino/blob/a839d4c87e6275851f8e08d518d1e0049d382adb/src/frontends/paddle/tests/test_models/gen_scripts/generate_quantize_linear.py
Is this the correct file?. I thought it would be a
c++ test file rather than python.

[xczhai]

@xczhai , Currently working on the errors.

If you have any questions, feel free reach me. I will reply asap.

https://github.com/xczhai/openvino/blob/a839d4c87e6275851f8e08d518d1e0049d382adb/src/frontends/paddle/tests/test_models/gen_scripts/generate_quantize_linear.py Is this the correct file?. I thought it would be a c++ test file rather than python.

For

@xczhai , Currently working on the errors.

If you have any questions, feel free reach me. I will reply asap.

https://github.com/xczhai/openvino/blob/a839d4c87e6275851f8e08d518d1e0049d382adb/src/frontends/paddle/tests/test_models/gen_scripts/generate_quantize_linear.py Is this the correct file?. I thought it would be a c++ test file rather than python.

@xczhai , Currently working on the errors.

If you have any questions, feel free reach me. I will reply asap.

https://github.com/xczhai/openvino/blob/a839d4c87e6275851f8e08d518d1e0049d382adb/src/frontends/paddle/tests/test_models/gen_scripts/generate_quantize_linear.py Is this the correct file?. I thought it would be a c++ test file rather than python.

Hi,

• The test case and paddle api may be out of date. I have attached the reference link in the tests. If the api is lagacy, you can refer to those paddle repo links. https://github.com/xczhai/openvino/blob/a839d4c87e6275851f8e08d518d1e0049d382adb/src/frontends/paddle/tests/test_models/gen_scripts/generate_quantize_linear.py#L27
• it is acceptable to cover the OP mapping correctness under some paddle model snippet for testing. C++ test is not needed here. You can refer to https://github.com/openvinotoolkit/openvino/tree/master/src/frontends/paddle/tests/test_models/gen_scripts.

[Pey-crypto]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

[xczhai]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

[Pey-crypto]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

[xczhai]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

[Pey-crypto]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

git submodule update --init --recursive --progress
cmake --build . --parallel $(nproc)
cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_MLAS=OFF -DENABLE_ONEDNN=OFF ..

[xczhai]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

git submodule update --init --recursive --progress cmake --build . --parallel $(nproc) cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_MLAS=OFF -DENABLE_ONEDNN=OFF ..

You should turn on mlas and onednn though it takes more time to build.

[Pey-crypto]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

git submodule update --init --recursive --progress cmake --build . --parallel $(nproc) cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_MLAS=OFF -DENABLE_ONEDNN=OFF ..

You should turn on mlas and onednn though it takes more time to build.

i did do that also, but then cmake reports mlas and onednn makefiles aren't available. I will try again and get back with the results.

[xczhai]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

git submodule update --init --recursive --progress cmake --build . --parallel $(nproc) cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_MLAS=OFF -DENABLE_ONEDNN=OFF ..

You should turn on mlas and onednn though it takes more time to build.

i did do that also, but then cmake reports mlas and onednn makefiles aren't available. I will try again and get back with the results.

Could you share the CPU platform and OS info? I can try to reproduce it locally.

[Pey-crypto]

@xczhai I am not able to build or test the commit I have made , because mlas and oneDNN are missing their MAKE files during the build. Using the -DENABLE flags doesn't seem to help. I have made two unit tests, for the functions and have not changed the underlying python api.

Did you execute git submodule update --init --recursive before building codes?

yup multiple times, i even cleared the submodules again and tried it again. I even used the --progress and --verbose flags to see the git clone process underneath, but all the clones were happening successfully

please share the commands you have tried.

git submodule update --init --recursive --progress cmake --build . --parallel $(nproc) cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_MLAS=OFF -DENABLE_ONEDNN=OFF ..

You should turn on mlas and onednn though it takes more time to build.

i did do that also, but then cmake reports mlas and onednn makefiles aren't available. I will try again and get back with the results.

Could you share the CPU platform and OS info? I can try to reproduce it locally.

Operating System: Ubuntu 24.04 LTS
Model name: 12th Gen Intel(R) Core(TM) i5-12500H
Architecture: x86-64

[Pey-crypto]

@xczhai Finally figured out the error, the permissions on those folders were messed up somehow, attaching strace revealed perms were being changed. I guess it was a local issue on my end.

[Pey-crypto]

Hi @Pey-crypto Thanks for your commits. But we can polish this PR for better maintain.

1. please check https://github.com/openvinotoolkit/openvino/blob/master/src/frontends/paddle/src/internal/pass/transform_fakequantize.cpp is needed or not. Theoretically, such pass should be remove since your commit works.

2. Previously, I prepare some unittests for Quantization. Maybe you can import these tests to such PR. It can ensure the motification accuracy. https://github.com/xczhai/openvino/tree/refactor_paddle_slim

BR.

Hi @xczhai,
Removed transform_fakquantize as the node is being made in quantize_linear. Removed references of fuse_fakequantize_ops from the paddle frontend.

[mlukasze]

@xczhai please evaluate current state of the PR, enable jenkins pipelines if good enough

[Pey-crypto]

@xczhai I used structured bindings, which wasn't compliant with C++ 11 standard, I have replaced it with tie.

[xczhai]

@xczhai I used structured bindings, which wasn't compliant with C++ 11 standard, I have replaced it with tie.

Yes, you are right. The current implementation should be aligned with C++11.

[Pey-crypto]

@xczhai , I have fixed the tests. Could you review it?

[xczhai]

@xczhai , I have fixed the tests. Could you review it?

OK. I start it.