build(deps): extend Python support to 3.14 and relax numpy to >=2,<3 - #189
Open
laserkelvin wants to merge 26 commits into
Open
build(deps): extend Python support to 3.14 and relax numpy to >=2,<3#189laserkelvin wants to merge 26 commits into
laserkelvin wants to merge 26 commits into
Conversation
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Gate the RAPIDS stack (nvidia-physicsnemo[cuXX] extra, cuml-cuXX) to python_version < '3.14' in both cu12 and cu13 extras: RAPIDS publishes no cp314/cp315 wheels (cuml-cu12 per-interpreter wheels; cuml-cu13 abi3 but its cupy-cuda13x dep lacks cp314), which otherwise blocks universal resolution with the 3.15 ceiling. Core physicsnemo remains on 3.14+. Resolution status: cu12/cu13 resolve on 3.11-3.14 (numpy pins to 2.2.6 under the 3.11 floor; 2.3+ needs >=3.12 upstream). 3.15 blocked by dev dep rdkit (cp314 max) until new wheels ship. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
…verride RAPIDS supports Python 3.14 from 26.4.0 via cp311-abi3 wheels (RSN 58), so the python_version gates were wrong. The real 3.14 blocker was nvidia-physicsnemo[cuXX]'s cupy-cuda1Xx<14 pin (cupy <14 has no cp314 wheel and no sdist); override it to >=14.0.1,!=14.1.0, matching cuml 26.8's own requirement. Both extras now resolve their full stack on 3.11-3.14; on 3.14 the cu12 fork picks torch 2.14.0+cu126. rdkit stays gated <3.15 (2026.3.6 ships cp315 only for aarch64/mac) and is unused in the codebase. 3.15 remains blocked upstream by tensordict (no cp315 wheel or sdist in any release), triton, and cupy. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
The 2.2.6 pin was lockfile stickiness from the old numpy<2.4 spec, not a constraint of the 3.11 floor. 2.4.6 supports >=3.11 with cp311-cp314 wheels, so 3.14 now gets a wheel instead of an sdist build. Full suite green on 3.13 and 3.14 with 2.4.6. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
fairchem-core pins lmdb<=1.7.3, which has no cp314 wheel. lmdb>=2.0.0 ships cp314 wheels and nothing in our fairchem usage touches lmdb, so override the pin from 3.14 on. Below 3.14, re-state fairchem's original pin explicitly: a marker-scoped override alone replaces the requirement outright, which would drop lmdb entirely on 3.13 and silently break the uma env. Verified in isolation on 3.14.7: fairchem imports, 33 UMA wrapper/equivalence tests pass with lmdb 2.3.0. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
hooks/reporting/layouts/base.py and training/cli.py call plt.clf(), removed in plotext 6. The lock kept 5.3.2 only by resolution stickiness; any fresh resolve (uv lock --upgrade, pip) would install 6.x and crash the Rich dashboard and training CLI. Found during the Python 3.15 exploration (plotext 6.1.0 broke the same paths there). Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Validated on 3.15.0rc1 (L4, cu126): core stack runs — ~3,554 tests and a GPU dynamics smoke pass with tensordict built from source, torch 2.14.0+cu126 cp315 wheels, and an eager torch.compile shim. CUDA extras stay 3.14-max until upstream ships cp315 wheels. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
…on315 Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com> # Conflicts: # pyproject.toml # uv.lock
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
15 tasks
Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com> # Conflicts: # CHANGELOG.md # pyproject.toml # uv.lock
laserkelvin
marked this pull request as ready for review
September 11, 2026 19:55
Collaborator
Author
|
/ok to test eb282e8 |
Contributor
 The PR appears safe to merge; no outstanding blocking or non-blocking code issue was established. Summary
Reviews (4) · Last reviewed commit: "build(deps)!: slim cu12/cu13 extras to P..." |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
dm-tree served a single call site (DataMixin.apply's map_structure over model_dump output) and nothing else in the lock required it; tensordict dropped it as a dependency years ago. The lock pinned 0.1.9, which has no cp314 wheel, so CI's 3.14 cell tried an sdist build that needs cmake — absent on the runners — and failed at the install step. Replacing the call with a small dict/list/tuple mapper removes the compiled dependency entirely (also a Python 3.15 blocker, no cp315 wheel exists). test/data: 823 passed each on 3.13 and 3.14. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
3.15 is not publicly supported yet: tensordict, cupy, and triton have no cp315 artifacts, so ordinary installs cannot resolve. Keep the published range honest at >=3.11,<3.15 (also avoids <=3.15's trap of excluding 3.15.1+ patch releases); internal 3.15 development continues via the recipe in docs/dev/python-315-internal.md. Addresses Greptile P1 on unsupported 3.15 metadata. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Collaborator
Author
|
/ok to test 53aafd7 |
…stack Nothing in nvalchemi imports cuml, cupy, pylibraft, or DALI; the RAPIDS stack entered via physicsnemo's [cuXX] extras and our duplicate cuml lines, whose upstream pins (cupy<14, pandas<2.4, numba caps) made 'pip install nvalchemi-toolkit[cuXX]' unresolvable on Python 3.14 and required uv-only cupy overrides. The extras now provide the CUDA torch build, toolkit-ops, cuequivariance-ops-torch, and physicsnemo core; users needing RAPIDS install it directly. Verified in an isolated 3.14 env: full import surface plus test/models + test/data green with and without RAPIDS; pure-pip resolution of the slimmed cu12 set succeeds. Lock: 382 -> 324 packages (numpy 2.5.3 main / 2.4.6 under fairchem's numpy<2.5 cap in the uma fork). Also aligns the install/UMA docs with the slimmed extras, documents the breaking change in the CHANGELOG, removes the accidentally committed docs/dev internal notes, and fences the pre-existing install-matrix HTML that trips MD013. BREAKING CHANGE: cu12/cu13 extras no longer include the RAPIDS stack. Signed-off-by: Kelvin Lee <kinlongkelvi@nvidia.com>
Collaborator
Author
|
/ok to test 707387d |
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
1 participant
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.
ALCHEMI Toolkit Pull Request
Description
Extends supported Python versions and relaxes numpy for
nvalchemi-toolkit:requires-pythongoes from>=3.11,<3.14to>=3.11,<3.15: Python 3.14 is now fully supported (core package, cu12/cu13 CUDA extras, uma/uma-cu12/uma-cu13 extras). Python 3.15 is deliberately not publicly supported yet (upstream wheels missing); internal 3.15 development is enabled by the recipe indocs/dev/python-315-internal.md.numpy<2.4becomesnumpy>=2,<3— no code changes needed; the lock resolves numpy 2.5.3 for the main envs (2.4.6 where fairchem'snumpy<2.5cap applies in the uma fork). Downstream users are free to use any numpy 2.x.cu12/cu13extras are slimmed to PhysicsNeMo core (CUDA torch +nvalchemi-toolkit-ops+cuequivariance-ops-torch+ physicsnemo core): the RAPIDS stack (cuml,cupy,pylibraft, DALI) is removed — nothing innvalchemiimports it, and its upstream pins made pure-pipinstalls unresolvable on 3.14. Users needing RAPIDS install it directly. This is the PR's one breaking change (install surface only), documented in the CHANGELOG.cuml-cuXX>=26.4.0, the cp311-abi3 line with official 3.14 support, RSN 58) and overridingnvidia-physicsnemo[cuXX]'scupy-cuda1Xx<14pin (cupy <14 has no cp314 wheel and no sdist).lmdboverride pair (fairchem pinslmdb<=1.7.3, which has no cp314 wheel; nothing in our fairchem usage touches lmdb). Below 3.14 fairchem's original pin is re-stated explicitly so the override doesn't drop the dependency.plotext<6: plotext 6 removedclf(), whichhooks/reportingandtraining/clicall — any fresh resolve would have installed 6.x and crashed the Rich dashboards and training CLI. The lockfile only kept 5.3.2 by resolution stickiness, so this was a latent breakage on every Python version.Type of Change
Related Issues
None — exploratory dependency-bounds work. Follow-ups tracked on the team board: nvalchemi-toolkit-ops metadata + lazy
torch.compile, cu13 runtime testing, cuequivariance-ops floor bump.Changes Made
requires-pythonto>=3.11,<3.15; relax numpy to>=2,<3cu12/cu13extras to PhysicsNeMo core; drop the RAPIDS stack and the (now-unnecessary) cupy overrides (breaking, CHANGELOG'd)lmdbper-Python override pair for fairchem on 3.14python_version < '3.15'(no cp315 x86_64 wheel; zero imports in the codebase)plotext<6(6.x removedclf(), used by hooks/reporting and training/cli)docs/dev/python-315-internal.md: validated internal 3.15 dev recipeTesting
make lintnot run locallyAlso verified by dry-run resolution on the merged state: cu12/cu13 extras resolve on 3.11–3.14;
umaresolves on 3.13 (lmdb 1.7.3, fairchem's pin preserved) and 3.14 (lmdb 2.3.0); the newuma-cu12extra resolves on 3.14 (fairchem 2.22.0 + lmdb 2.3.0 + torch 2.13.0+cu126).Checklist
docs/dev/python-315-internal.mdAdditional Notes
pipinstalls now work on 3.11–3.14: the RAPIDS removal (and physicsnemo core instead ofphysicsnemo[cuXX]) eliminates thecupy<14pin that madepip install nvalchemi-toolkit[cu12]unresolvable on 3.14 — verified: pip resolves the slimmed cu12 set on 3.14 in ~90s.[tool.uv] override-dependenciesentries that remain (lmdb) are uv-only and only tighten uv's lock; pip users are unaffected.nvalchemi-toolkit-opsstill declaresrequires-python <3.15(metadata-only cap; a sibling release is queued to relax it); CUDA extras cannot install on 3.15 until upstream (cupy/RAPIDS/cuequivariance) ships cp315 wheels.