Vector Assertion Failure in InferenceSession Init with Hotplugged-Off Cores on ARM (v1.21.0)

[PapperYZ]

Describe the issue

Environment:

• Device: OrangePi 5 Plus (4x A76 + 4x A55)
• Config: Any cores disabled (e.g., echo 0 > /sys/devices/system/cpu/cpu[4-7]/online)
• OS: Ubuntu 22.04
• ONNX Runtime: v1.21.0
• Python: 3.10
• GCC: 14 (gcc-toolset-14)

Issue:

• Crashes during InferenceSession init with
  /opt/rh/gcc-toolset-14/root/usr/include/c++/14/bits/stl_vector.h:1130: Assertion '__n < this->size()' failed.
• Succeeds when all 8 cores are online (even if restricted to 4 via taskset -c 4-7) or on a 4x A53 system with no hotplugging.
• Fails whenever any cores are hotplugged off (e.g., 4x A55 only or A76 + partial A55).

Observations:

• os.cpu_count() correctly reports online cores (e.g., 4).
• OMP_NUM_THREADS=1 doesn’t fix it—issue is in init, not runtime threading.
• Suspect: Mishandles CPU topology when cores are hotplugged off (e.g., uses present instead of online).

To reproduce

disable 4 cores with command below

echo 1 | sudo tee /sys/devices/system/cpu/cpu[0-3]/online
echo 0 | sudo tee /sys/devices/system/cpu/cpu[4-7]/online

and run simple python script below:

import onnxruntime
model_path = "/path/to/resnet50_v1.onnx"
session = onnxruntime.InferenceSession(model_path, providers=["CPUExecutionProvider"])
print("Session initialized successfully")

You will see:

Urgency

please help to solve the issue soon, it should be a quick fix... it is currently preventing us with a benchmark reference with our ASIC with similar cores.

Platform

Linux

OS Version

Ubuntu 22.04

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

ONNX Runtime: v1.21.0

ONNX Runtime API

Python

Architecture

ARM64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[snnn]

Could you please build ONNX Runtime in debug mode and generate a stacktrace?
Doc: https://onnxruntime.ai/docs/build/inferencing.html

[PapperYZ]

Sorry, I do not know how to do that
I have already provided clear reproduce steps, would you try to look into that?