Failed to load model in virtual machine.

[fjea]

Hi guys:

I tried and failed to load a trained model using TensorFlow in VIRTUAL MACHINE with Linux OS , but it works well in physical machine with the same Linux OS.
Any help would be appreciated,thanks.

The specific details are as follows:

the dupmed err :
hs_err_pid3013.log

==================
jvm 2 | 2026-08-19 16:59:19.715 [WrapperSimpleAppMain] INFO s.f.e.a.c.EqNoiseClassification.loadModel:61 - Building model loading standard...
jvm 2 | 2026-08-19 16:59:19.735 [WrapperSimpleAppMain] INFO s.f.e.a.c.EqNoiseClassification.loadModel:73 - Model loading standard built successfully
jvm 2 | 2026-08-19 16:59:19.735 [WrapperSimpleAppMain] INFO s.f.e.a.c.EqNoiseClassification.loadModel:75 - Model loading...
Loading: 100% |████████████████████████████████████████|
jvm 2 | 2026-08-19 16:59:23.794 [WrapperSimpleAppMain] INFO a.d.u.Platform.detectPlatform:84 - Found matching platform from: jar:file:/home/ndws-eew-channel-1.0.2-SNAPSHOT/lib/tensorflow-native-cpu-2.16.1-linux-x86_64.jar!/native/lib/tensorflow.properties
jvm 2 | #
jvm 2 | # A fatal error has been detected by the Java Runtime Environment:
jvm 2 | #
jvm 2 | # SIGILL (0x4) at pc=0x00007fbae62a154a, pid=3013, tid=3039
jvm 2 | #
jvm 2 | # JRE version: OpenJDK Runtime Environment BiSheng (17.0.19+13) (build 17.0.19+13)
jvm 2 | # Java VM: OpenJDK 64-Bit Server VM BiSheng (17.0.19+13, mixed mode, sharing, tiered, compressed class ptrs, g1 gc, linux-amd64)
jvm 2 | # Problematic frame:
jvm 2 | # C [libtensorflow_framework.so.2+0x153454a] google::protobuf::internal::OnShutdownRun(void ()(void const), void const*)+0xba
jvm 2 | #
jvm 2 | # Core dump will be written. Default location: Core dumps may be processed with "/usr/lib/systemd/systemd-coredump %P %u %g %s %t %c %h" (or dumping to /home/ndws-eew-channel-1.0.2-SNAPSHOT/bin/core.3013)
jvm 2 | #
jvm 2 | # An error report file with more information is saved as:
jvm 2 | # /home/ndws-eew-channel-1.0.2-SNAPSHOT/bin/hs_err_pid3013.log
jvm 2 | #
jvm 2 | # If you would like to submit a bug report, please visit:
jvm 2 | # https://gitcode.com/openeuler/bishengjdk-17/issues/
jvm 2 | # The crash happened outside the Java Virtual Machine in native code.
jvm 2 | # See problematic frame for where to report the bug.
jvm 2 | #

[0xshubhs]

Not a DJL problem. Your VM's CPU is missing an instruction set TensorFlow needs.

#  SIGILL (0x4) at pc=0x00007fbae62a154a, pid=3013, tid=3039
#  C  [libtensorflow_framework.so.2+0x153454a]  google::protobuf::internal::OnShutdownRun(...)

SIGILL is illegal instruction, the CPU got an opcode it doesn't implement. TensorFlow's official x86_64 Linux builds have required AVX since 1.6 and tensorflow-native-cpu-2.16.1-linux-x86_64 is one of them. Your physical box has AVX, the VM is presenting a generic CPU model that masks it.

OnShutdownRun is protobuf static init, so it dies during library load before your code runs. Matches your log, platform detected then gone.

Check on both machines:

grep -o -m1 'avx[0-9a-z_]*' /proc/cpuinfo | sort -u

Physical will show avx, avx2, fma. The VM won't.

Fix depends on hypervisor, all of them mean stop hiding host CPU features:

KVM/libvirt: <cpu mode='host-passthrough'/> or host-model
Proxmox: CPU type host instead of kvm64/x86-64-v2
VMware: check EVC isn't set below Sandy Bridge
Hyper-V: turn off processor compatibility mode

Needs a guest reboot, flags are fixed at boot.

If you can't change the CPU model there's no official non-AVX TF build. PyTorch and ONNX Runtime have lower baselines and would run.

[fjea]

Thank you so much, that really helps me.
As you said, "Your physical box has AVX, the VM is presenting a generic CPU model that masks it." my physical host shows avx, avx2, fma and the VM won't.
I changed the VM`S CPU work mode from “Compatibility Mode“ to “Host Matching Mode“， it shows AVXs , and then my trained model could be loaded and works well.
Thanks again！