Merge main into release: torch 2.10 / CUDA 13 / py3.12 + transformers v5 security floor - #397
Merged
Conversation
* build: migrate to torch 2.10 / CUDA 13 / Python 3.12 - Dockerfile: base image -> pytorch/pytorch:2.10.0-cuda13.0-cudnn9-runtime - pyproject: torch==2.10.0 from the cu130 index (tool.uv.index + sources), requires-python >=3.12, mamba-ssm/causal-conv1d bumped to cu13 wheels, drop loompy/scib/louvain/gitpython, move protobuf floor to constraint-dependencies, add pypi.nvidia.com index for cu13 deps Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XJGBowF2ixDugsVKn6LkQT (cherry picked from commit b4eedc1) * ci: pin Python 3.12.13; consolidate dev deps into [dev] extra - Bump all CI python-version pins (main.yml, release.yml) 3.11.x -> 3.12.13, matching requires-python >=3.12 from the torch 2.10 migration. - Move requirements-dev.txt into pyproject.toml as a `dev` optional-dependency and delete the file; jobs that run pytest now install .[mamba-ssm,dev] (also fixes the notebooks jobs that ran pytest --nbmake without installing it). - requirements_cuda.txt cannot move into pyproject (helical is published to PyPI, which rejects direct-URL deps even in extras), so refresh its stale cu12/torch2.7/cp311 wheels to the migrated cu13/torch2.10/cp312 set and document the constraint. - README: dev-install section, prebuilt-cuda-wheels path, and refresh stale conda 3.11 / cu128 / old mamba .whl install snippets. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRLET4TF6sECaBRwKvCbPb (cherry picked from commit b255cee) * ci: install accelerators from prebuilt wheels, not source build pip install .[mamba-ssm] fails on the fresh 3.12 runners: mamba-ssm and causal-conv1d ship only sdists on PyPI, their setup.py imports torch at the top level, and PEP 517 build isolation runs that build in an env without torch (torch is intentionally absent from their build-system.requires), so metadata generation dies with "No module named 'torch'". Install helical first (brings torch 2.10 — the default PyPI Linux wheel is the cu13 build, matching our accelerator wheels), then install the prebuilt mamba-ssm/causal-conv1d/flash-attn wheels from requirements_cuda.txt (no compile, no build isolation). Replaces the source-build .[mamba-ssm] path and the `flash-attn --no-build-isolation` step across every CI job. The .[mamba-ssm] extra stays in pyproject.toml for library users. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRLET4TF6sECaBRwKvCbPb (cherry picked from commit c9ea76f) * ci: install into a fresh uv venv per job (isolate from persistent runner env) The self-hosted runner reuses the setup-python tool-cache interpreter across runs, so `pip install` accumulated packages in a shared site-packages. After the cu12->cu13 torch migration a stale nvidia-nccl-cu12 (NCCL < 2.29.7) lingered and was loaded instead of the cu13 NCCL, so `import torch` died with `undefined symbol: ncclCommResume` (a userspace NCCL mismatch, not a driver issue). Create a fresh `uv venv` in $RUNNER_TEMP for each GPU/test job and install into it (exposed to later steps via $GITHUB_PATH / $GITHUB_ENV), so no stale packages can carry over between runs. The ubuntu fresh-install smoke test and the PyPI publish job stay on plain pip. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRLET4TF6sECaBRwKvCbPb (cherry picked from commit 7375e23) * ci: use astral-sh/setup-uv action instead of bootstrapping uv via pip Provide uv through the pinned astral-sh/setup-uv@v8.3.2 action (same as the other repos) rather than `python -m pip install uv` inside the run block — cached and consistent. The fresh uv venv per job is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRLET4TF6sECaBRwKvCbPb (cherry picked from commit 371c13b) * fix(deps): declare scanpy/requests-cache/torchmetrics/catalogue These are imported at module top level in helical but were never declared as dependencies — they were satisfied transitively (scanpy came via scib, removed in this migration) or by the persistent runner env. The fresh uv venv exposed them as collection-time ImportErrors (e.g. c2s: "No module named 'scanpy'"). - scanpy: scgpt / c2s / uce / tahoe model code - requests-cache: utils/downloader.py (weight download, hit by every model) - torchmetrics + catalogue: tahoe backbone evo2's Bio/vortex imports stay undeclared — they're guarded by try/except in the package __init__ and the test skips when absent, so they don't break collection. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01RRLET4TF6sECaBRwKvCbPb (cherry picked from commit cdbe557) * fix(notebooks): load NT downstream tasks via data_dir for datasets 3.x The migration bumps `datasets` 2.14.7 -> 3.6.0. The `InstaDeepAI/nucleotide_transformer_downstream_tasks_revised` dataset no longer ships a loading script that defines per-task BuilderConfigs; it is now plain parquet organized one directory per task, with no named configs declared. Under datasets 3.x the old positional-config call therefore fails: ValueError: BuilderConfig 'promoter_tata' not found. Available: ['default'] which broke the `notebooks` CI job (Hyena-DNA-Inference and HyenaDNA-Fine-Tuning) while all other jobs passed. Switch both notebooks to select the task via `data_dir=label`, which loads the correct train (5062) / test (212) promoter_tata splits with the same `sequence`/`label` columns the downstream cells use. Also print `label` instead of `.config_name` (now "default") in the inference notebook. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CpgHKFsoeKfqrmrFdvESX4 (cherry picked from commit a277bee) * chore(deps): bump pytest tooling for py3.12 migration Bump pytest 8.2.0 -> 9.0.3, pytest-cov 5.0.0 -> 7.1.0, and pytest-mock 3.14.0 -> 3.15.1 to versions compatible with the Python 3.12 / torch 2.10 target stack. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CpgHKFsoeKfqrmrFdvESX4 (cherry picked from commit 8093274) * release: bump version to 3.0.0 (major) This release drops previously supported runtimes, which is a backward-incompatible change under SemVer and warrants a major bump: - requires-python raised 3.11 -> 3.12 (Python 3.11 support dropped) - torch pinned 2.7.0 -> 2.10.0 (exact), forcing a core shared runtime bump - CUDA 12 -> CUDA 13 (cu13 wheels for torch/mamba/causal-conv1d/flash-attn); drops CUDA 12 GPUs / drivers < 580 - removed dependencies: loompy, scib, louvain, gitpython Existing users on Python 3.11 or CUDA 12 cannot upgrade in place, so this is not a patch- or minor-compatible change. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CpgHKFsoeKfqrmrFdvESX4 * fix(tests): remove no-op skipif marks on evo2 fixtures pytest 9 promotes "marks applied to fixtures" from a deprecation warning to a hard collection error, aborting the entire test session. The @pytest.mark.skipif on the read_prompts, evo2_model, and test_forward_pass fixtures never had any effect (marks on fixtures are ignored); the skipif on the test_evo2 test function still handles skipping when Evo 2 is absent. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EJk4wn3VJ4MNLgt6WzaFnj * Prepare transformers v5 compatibility + security floor Bump transformers to >=5.3.0 (clears CVE-2026-4372 / CVE-2026-1839) and make the affected code forward-compatible with the v4->v5 major bump: - grouped_query_attention: drop removed `position_ids` kwarg from apply_rotary_pos_emb (no-op on 4.x, required on 5.x) - hyena_dna: import AddedToken/PreTrainedTokenizer from the public transformers namespace (internal tokenization_utils path removed in v5) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JKLMeVCKGFNvCUpLugKuFK --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
bputzeys
approved these changes
Jul 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Promote the torch 2.10 / CUDA 13 / Python 3.12 migration and transformers>=5.3.0 security floor (PR #394) from main to release.
🤖 Generated with Claude Code