update docker image for segmentation method 'hmm'

[kmavrommatis]

Using the cnvkit docker image cnvkit:0.9.13
In order to run with segmentation method 'hmm' the package pomegranate is required.
Furthermore, the torch version needs to be updated from the current version 2.7.

In order for it to work I had to update the docker image with the commands:

pip uninstall -y torch pomegranate && \
pip install pomegranate>=1.0.0

Thanks

[etal]

Sorry for the oversight. This module's dependencies have been quite the journey, so in short you won't need pomegranate or torch for CNVkit 0.9.14 which I hope to release in the next couple of weeks.

[etal]

Both a forward fix and a backport for the 0.9.13 image are now available.

Root cause. The problem was not that pomegranate was missing — it is declared as a dependency — but that the torch it depends on could not be imported. The 0.9.13 image installed pomegranate from conda-forge, which pulled in the libtorch 2.7.0 (cpu_openblas) build. That build ships libtorch_cpu.so with two defects: an executable-stack ELF flag that modern loaders refuse (cannot enable executable stack as shared object requires), and a protobuf/abseil ABI mismatch (undefined symbol: absl...MutexD1Ev). Either one breaks import torch, which makes pomegranate, and therefore HMM segmentation, unavailable. Your workaround worked because reinstalling from PyPI replaced the broken conda torch with a self-contained manylinux wheel that has neither defect.

Forward fix — CNVkit 0.9.14 (recommended). HMM segmentation has been reimplemented directly on SciPy; pomegranate and torch are no longer dependencies at all. The 0.9.14 image (etal/cnvkit:0.9.14, also tagged latest) therefore removes the entire problematic dependency chain. Upgrading is the durable fix.

Backport — the etal/cnvkit:0.9.13 image. For anyone pinned to the 0.9.13 line, the 0.9.13 image tag has been rebuilt and republished with a working environment. The CNVkit code is unchanged (this is an image-only fix): pomegranate is no longer installed via conda, and the CPU-only PyTorch wheel is installed from PyPI instead, so import torch and HMM segmentation now work out of the box. The republished image is verified end-to-end (cnvkit.py segment -m hmm runs successfully), and the change did not move the latest or 0.9 tags.

docker pull etal/cnvkit:0.9.13 (or :0.9.14) will fetch the corrected image; if you have the old one cached, pull again to pick up the new digest.

Thanks for the clear report and the reproduction steps.