algorithm-nexus[product]==0.2.2 broken on fresh pip install: torchgeo 0.10.0 incompatibility

[tmuttaki]

Summary

A fresh pip install algorithm-nexus[product]==0.2.2 fails at runtime because pip resolves torchgeo==0.10.0, which is incompatible with the pinned terratorch[vllm]==1.2.10.

Error

ImportError: cannot import name 'utils' from 'torchgeo.trainers'
(/opt/vllm/lib64/python3.12/site-packages/torchgeo/trainers/__init__.py)

This breaks both the terratorch_fix vLLM plugin and the Terratorch model architecture registration, making geospatial models (e.g. ibm-nasa-geospatial/Prithvi-EO-2.0-300M-TL-Sen1Floods11) unservable via vLLM.

Root Cause

1. torchgeo 0.10.0 (released 2026-08-14) renamed torchgeo.trainers → torchgeo.tasks (torchgeo PR #3887). The trainers.utils submodule was not shimmed with a deprecation alias — it was simply removed.

2. terratorch 1.2.10 (the version pinned by algorithm-nexus) imports from torchgeo.trainers import utils in multiple files (timm_model_factory.py, cli_tools.py, base_task.py, reconstruction_tasks.py, wxc_downscaling_task.py). Its pyproject.toml declares torchgeo as a dependency with no upper bound.

3. algorithm-nexus 0.2.2 pins terratorch[vllm]==1.2.10 in the [product] extra but has no direct torchgeo version constraint. The requirements-product.txt lock file pins torchgeo==0.9.0, but a bare pip install algorithm-nexus[product]==0.2.2 (without the lock file) lets pip resolve torchgeo==0.10.0, which breaks the import.

Upstream Status

• terratorch PR #1222 (merged 2026-08-18 to main) pins torchgeo>=0.9.0,<0.10 — but this fix is not in any release. Latest release is still v1.2.10 (2026-07-15), which predates the fix.
• terratorch PR #1221 (open) does a full API migration to support torchgeo 0.10.

Reproduction

# In a clean environment or container (e.g. quay.io/vllm/automation-vllm:cuda-*)
pip install algorithm-nexus[product]==0.2.2

# This fails:
python -c "from terratorch.vllm.plugins.general import register_terratorch_fix"
# ImportError: cannot import name 'utils' from 'torchgeo.trainers'

# Workaround:
pip install "torchgeo<0.10"
# Now the import succeeds

[tmuttaki]

Additional Issue: vLLM version downgrade

algorithm-nexus[product]==0.2.2 pins vllm==0.26.0 as an exact dependency. When installed into an RHAIIS container image that ships a midstream vLLM build (e.g. vllm 0.26.1.dev54+rhaiv.3), pip downgrades the entire vLLM package to upstream 0.26.0.

This overwrites all midstream patches, security fixes, and RHAIIS-specific changes (including the built-in Terratorch model class, Dockerfile fixes from nm-vllm-ent PR #658, etc.).

The [product] extra in pyproject.toml:

product = [
    "opencv-python-headless!=5.0.0.93",
    "terrakit==0.2.0",
    "terratorch[vllm]==1.2.10",
    "vllm==0.26.0",               # forces downgrade of midstream vllm
]

If algorithm-nexus is intended to be installable into pre-built RHAIIS images, the vllm==0.26.0 pin should be loosened (e.g. vllm>=0.26.0,<0.27) or removed entirely so it uses whatever vLLM the image already provides.

[christian-pinto]

Thanks for the heads up. We are working with the Terratorch team to fix the torchgeo issue. Regarding, vLLM we are pinning it to whatver version is targeted for the product release. Also, the exact pin guarantees that the models installed via nexus have been tested with and are installed alongside a version of vLLM that is compatible.