[Bug]: CPU plugin compilation drops ReduceMax operations, changes output shape

[hari-krishna2000]

OpenVINO Version

2025.3.0

Operating System

Other (Please specify in description)

Device used for inference

CPU

Framework

PyTorch

Model used

Custom TorchScript model converted to OpenVINO IR format

Issue description

OS: Ubuntu 22.04 LTS

After calling compile_model() with CPU device, the output shape of scalar outputs incorrectly changes from [1] to [1, 960]. Investigation using get_runtime_model() reveals that all 21 ReduceMax operations are dropped during CPU plugin compilation. The same holds for the GPU model.

The model's IR graph is correct - it shows ReduceMax_17: [1,960] -> [1] producing a scalar. But after CPU compilation, this reduction is missing and the output retains the intermediate shape [1, 960].

The model was converted from TorchScript using OpenVINO's PyTorch frontend.

Step-by-step reproduction

import openvino as ov

core = ov.Core()
model = core.read_model("model.xml")
model.reshape(input_size)

print("Before compile:")
print(len(model.get_ops()))
print(len([op for op in model.get_ops() if op.get_type_name() == 'ReduceMax'])})
print(model.outputs[0].get_partial_shape()})

compiled_model = core.compile_model(model, "CPU")
print(compiled_model.outputs[0].get_partial_shape()})

The latter shows up as 0; and the shape is changed.

Relevant log output

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.

[ShreyasN707]

Hi! @hari-krishna2000 I am new here and I am interested in this issue. May I work on this issue?

[hari-krishna2000]

@ShreyasN707 feel free to!

[ShreyasN707]

@ShreyasN707 feel free to!

Sure! please assign it to me.

[Hmm-1224]

Hi @hari-krishna2000,

Thanks for reporting this issue. From your description, it seems that the CPU plugin is incorrectly dropping ReduceMax operations during compilation, which is causing the output shape to change.

A few approaches we could try:

1. Insert an identity operation after each ReduceMax to prevent the optimizer from removing it.
2. Ensure your model input is statically shaped before compilation, as dynamic axes can trigger ReduceMax removal.

Also, there are minor syntax issues in your code, please correct them to avoid runtime errors.

[Hmm-1224]

.take

[Hmm-1224]

@hari-krishna2000,
I tested this on my machine running Ubuntu 24.04.1 LTS. I created a PyTorch model with two ReduceMax operations, converted it to OpenVINO IR, compiled it on CPU, and checked the ReduceMax operations before and after compilation. Everything worked fine, the ReduceMax operations were retained and the output shapes remained correct. I’m not sure if this behavior is specific to Ubuntu 22.04 LTS or the OpenVINO version, so it would be great if ou could try on that setup and share the results.