Migrate model/metadata loading from wandb to HuggingFace

[avantikalal]

Summary

• Replaces broken grelu.resources.get_artifact() / wandb-based loading with hf_hub_download from HuggingFace Hub
• Model weights: Genentech/decima-model, metadata: Genentech/decima-data
• Preserves full wandb support for internal/private models in new decima.hub.wandb submodule (uses wandb.Api() directly, no gReLU dependency)
• Drops upper bound on gReLU pin (>=1.0.10)

Changes

File	Change
src/decima/constants.py	Add HF_MODEL_REPO, HF_DATA_REPO, HF_METADATA_FILENAME
src/decima/hub/__init__.py	Rewrite: HuggingFace primary backend, local paths preserved
src/decima/hub/wandb.py	New: all wandb functions via wandb.Api() directly
src/decima/hub/download.py	Update: hf_hub_download with local_dir
setup.cfg	Add huggingface_hub dep, add wandb/hf pytest markers
tests/test_hub.py	Update to HF; mark model/metadata tests @pytest.mark.hf
tests/conftest.py	Remove login_wandb() call (HF is public)

Internal user migration

# Before
from decima.hub import load_decima_model

# After (for private/internal wandb models)
from decima.hub.wandb import load_decima_model
model = load_decima_model("private_model", host="https://internal.wandb.company.com")

Test plan

• test_hf_constants — constants correct
• 39 tests pass with pytest -m "not wandb and not hf and not long_running"
• test_load_result passes (downloads metadata.h5ad from HF, ~3.3 GB)
• @pytest.mark.hf tests (full model download) — require HF network access
• @pytest.mark.wandb tests — require wandb credentials

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR migrates Decima’s public model/metadata loading from the previously broken wandb/gReLU artifact mechanism to the HuggingFace Hub, while keeping wandb-based loading available for private/internal usage via a new decima.hub.wandb module.

Changes:

• Switched decima.hub and decima.hub.download to use huggingface_hub.hf_hub_download for model weights and metadata.
• Added HF Hub constants (HF_MODEL_REPO, HF_DATA_REPO, HF_METADATA_FILENAME) and the huggingface_hub dependency.
• Updated tests and CI configuration to introduce hf/wandb markers (but currently leaves HF tests enabled by default).

Reviewed changes

Copilot reviewed 9 out of 10 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/decima/constants.py	Adds HF Hub repo/filename constants used by the new loading path.
src/decima/hub/__init__.py	Replaces wandb artifact loading with HF Hub downloads for model + metadata.
src/decima/hub/download.py	Updates download/cache utilities to fetch weights/metadata from HF Hub.
src/decima/hub/wandb.py	Introduces wandb-only loader utilities using wandb.Api() for private/internal usage.
tests/test_hub.py	Moves hub tests to HF and marks download tests with @pytest.mark.hf.
tests/conftest.py	Removes unconditional wandb login from test startup.
setup.cfg	Adds huggingface_hub, adjusts gReLU pin, and defines pytest markers/default marker filter.
docs/tutorials/5-gene-expression-prediction.ipynb	Updates tutorial outputs/content consistent with HF-based loading.
.github/workflows/run-tests.yml	Tweaks CI tox install/run invocation (but currently still runs HF tests by default).

Comments suppressed due to low confidence (2)

tests/test_hub.py:28

• These tests are marked hf but not long_running, and they download full model weights/metadata. If hf is ever included in default runs (or a user forgets to exclude it), this will be extremely slow and bandwidth-heavy. Consider additionally marking them long_running (so they are skipped unless explicitly enabled) or otherwise gating them behind an opt-in flag/env var.

@pytest.mark.hf
def test_load_decima_model():
    model_0 = load_decima_model()
    assert model_0 is not None
    assert isinstance(model_0, LightningModel)

    model_2 = load_decima_model(model=2)
    assert model_2 is not None


@pytest.mark.hf
def test_load_decima_metadata():
    metadata = load_decima_metadata()
    assert isinstance(metadata, anndata.AnnData)
    assert metadata.shape == (8856, 18457)

docs/tutorials/5-gene-expression-prediction.ipynb:2223

• This tutorial cell shows a failing command (../tests/data/seqs.fasta not found). Either update the path to a file that exists in the repo (or add the referenced file), or replace the cell with an embedded example so readers can follow the tutorial without hitting a file-not-found error.

   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "cat: ../tests/data/seqs.fasta: No such file or directory\n"
     ]
    }
   ],
   "source": [
    "! cat ../tests/data/seqs.fasta | cut -c 1-200"
   ]

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.