🧬 🗺️ GLMap scoring containers

Prebuilt Apptainer / Singularity images that carry the GPU runtime environments for scoring all 123 genomic language models (gLMs) profiled in GLMap — Profiling genomic language models as individuals in a population.

The 123 models span runtime stacks that are mutually incompatible (Python 3.8–3.12, PyTorch 1.13–2.9, CUDA 11.7–12.4) and can never share one interpreter. These four images package every environment so you can recompute the likelihood responses without setting up a single conda env.

Only need the analysis (precomputed scores, figures/tables)? You don't need these images at all — pip install -e . on the GLMap repo gives a torch-free stack.

What's here

Each image is self-contained (the shared CUDA 12.8 base is already inside; download only the group(s) for the models you want to score):

Each image holds its envs as isolated micromamba environments and dispatches to the right one per model via the GLMAP_ENV variable.

Download

# one image
hf download Tim419/GLMap-containers bio-default.sif --repo-type dataset --local-dir .

# or all four
hf download Tim419/GLMap-containers --repo-type dataset --local-dir .

Run

Bind your GLMap checkout at /work (code, panel, audit, model weights) and pick the env with GLMAP_ENV:

GLMAP_ENV=caduceus apptainer run --nv \
    --bind "$PWD":/work --pwd /work bio-cu118.sif \
    scripts/score/scoring_worker.py --from-audit \
    --hf-ids kuleshov-group/caduceus-ph_seqlen-131k_d_model-256_n_layer-16

Or run the full 123-model sweep straight through the images:

python scripts/score/run_scoring_sweep.py \
    --backend container --image-dir <dir with the .sif files> --hf-cache "$HF_HOME"

• --nv exposes the host GPU.
• On compute nodes without user namespaces, use singularity run --nv (the same .sif) — the GLMap sweep takes --container-runtime singularity.
• HyenaDNA / megaDNA also need their loader code on the bound checkout: it is in the GLMap repo after bash models/setup_external_models.sh (HyenaDNA's is vendored; megaDNA's weight auto-downloads from the HF Hub).

See container/README.md and models/env_routing.md for the full model → image/env routing.

License

These images bundle many third-party open-source runtimes (PyTorch, Transformers, mamba-ssm, flash-attn, evo2, …) and each model family's loader code — each remains under its own upstream license. The GLMap glue code is Apache-2.0. Individual model weights are downloaded separately and follow their own licenses. Consult each upstream project before redistribution or commercial use.

Citation

@article{hou2026glmap,
  title   = {Profiling genomic language models as individuals in a population},
  author  = {Hou, Yusen and Long, Weicai and Su, Houcheng and Feng, Junning and Zhang, Yanlin},
  journal = {In submission},
  year    = {2026}
}

Project: https://github.com/ai4nucleome/GLMap · Panel dataset: Tim419/GLMap-panels