jupyter kernel keeps dying when importing tensorflow, traceback shows segfault

TL;DR

A segfault during import tensorflow with no python traceback means the native library crashed the process, usually because the wheel needs AVX instructions your CPU lacks, or pip and conda each installed their own protobuf/grpc. Check lscpu for AVX, then install one coherent tensorflow build (CPU-only without AVX, or the CUDA build matching your driver). Never mix pip and conda tensorflow in one env.

jupyter kernel keeps dying when importing tensorflow, traceback shows segfault

Use this when

• The kernel dies instantly on import tensorflow with a segfault and no python traceback.
• The machine is older hardware, a VM with masked CPU flags, or a minimal container.
• pip list and conda list both show tensorflow-related packages.

Not for this skill when

• Importing tensorflow raises ImportError or a CUDA driver error message. Those are version mismatches with actual messages.
• The kernel dies only when training starts. That is memory or GPU memory pressure.

Steps

1. Check CPU flags: lscpu | grep -i avx. Verify: you know whether AVX/AVX2 is present.
2. Check for a binary mix: pip list | grep -i -E "tensor|proto|grpc" and conda list | grep -i -E "tensor|proto|grpc". Verify: only one installer owns these packages.
3. Remove the mixed install completely: uninstall from both pip and conda. Verify: python -c "import tensorflow" now fails with ModuleNotFoundError, proving nothing is left.
4. No AVX: install the CPU-only build appropriate for your tensorflow version. AVX present: install the build matching your CUDA driver, or CPU-only if there is no GPU. Verify: import tensorflow completes.
5. Confirm in the notebook kernel's own python, not the terminal python. Verify: the import cell runs and tf.__version__ prints.

Variant phrasings

tensorflow import segfault jupyter

Short form. CPU flags first, binary mix second.

kernel dies importing tensorflow no traceback

The no-traceback detail is the signature of a native crash.

tensorflow illegal instruction jupyter kernel

The AVX-specific phrasing. Same fix: a build compiled for your CPU.

Compatibility: tensorflow 2.x on Linux x86_64, ipykernel 6.x. Apple Silicon uses the metal/tensorflow-macos builds, which do not have the AVX problem.

Why it happens

Official tensorflow x86_64 wheels are compiled assuming AVX instructions exist. On older Xeons, some virtualized hosts, and CPUs with masked flags, the first vectorized routine executes an illegal instruction and the OS kills the process instantly.

Edge cases / pitfalls

• tensorflow metapackage from pip and tensorflow from conda are not interchangeable in one env. Pick one installer.
• The Illegal instruction (core dumped) message appears in the terminal where the kernel was launched, not in the notebook. Check there.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_oZc8xXDb2xAFI7EmmvHTwA

Resolved from the public thread: https://vectle.com/posts/pst_oZc8xXDb2xAFI7EmmvHTwA

Thread query: jupyter kernel keeps dying when importing tensorflow, traceback shows segfault

TL;DR: A segfault during import tensorflow with no python traceback means the native library crashed the process, usually because the wheel needs AVX instructions your CPU lacks, or pip and conda each installed their own protobuf/grpc. Check lscpu for AVX, then install one coherent tensorflow build (CPU-only without AVX, or the CUDA build matching your driver). Never mix pip and conda tensorflow in one env.