Skip to content

v1.3.1

Choose a tag to compare

View all tags
@IanShi1996 IanShi1996 released this 19 Oct 05:31
· 1 commit to main since this release
v1.3.1
582798e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Note

Superseded by v1.3.2, which fixes mRNABERT failing to load.

Patch release fixing the multimolecule-backed models, which were broken in v1.3.0 and recent v1.2.x installs. Upgrade if you use RiNALMo, RNABERT, RNAErnie, RNA-MSM, 3UTRBERT, UTR-LM or ERNIE-RNA.

Fixes

  • Random weights. The multimolecule checkpoints on HuggingFace now store weights under a model. prefix that multimolecule 0.0.7 doesn't recognize, so no weights were loaded. Each model now loads its checkpoint explicitly and raises if any parameter is missing, so this can't fail silently again.
  • ERNIE-RNA. Fixed a tokenizer vocab mismatch (26 vs 28) and renamed pairwise projection layers.
  • Import crash. Pinned danling==0.3.13. multimolecule leaves it unpinned, and newer versions need torch>=2.4.
  • Install. The README's base_models install now works: 
    pip install torch==2.2.2 --index-url https://download.pytorch.org/whl/cu121
pip install psutil
pip install --no-build-isolation mrna-bench[base_models]

Verification

  • Fresh-env install and full test suite pass for all seven models (the only failure is an Evo1 test that runs out of memory on a 24 GB GPU).
  • RiNALMo-giga matches the official reference value.
  • Hidden states for all 14 checkpoints were cross-checked against independently re-hosted copies of the models.

Full changelog: v1.3.0...v1.3.1 (#28)

Assets 2
Loading