Repository navigation
feat(uma): move to fairchem-core 2.25.0 and keep Nitro out of reach - #97
Conversation
Tracking: #61 (model-support spec §6.4) What was wrong: the pinned fairchem-core 2.21.0 cannot load the UMA-S 1.2.1 checkpoints (uma-s-1p2p1*.pt); they need >= 2.22. The runtime bump is also the first release whose upstream default changed semantics: 2.22+ made UMA Nitro the upstream "default" preset (Ampere and newer, reduced precision) and removed "turbo", so simply re-pointing the pin would have silently changed the precision of every mliport "default" run and could have selected Nitro. What changes: - fairchem-core 2.25.0 (latest, Requires-Python >=3.11,<3.15) on torch 2.13.0. 2.24 and 2.25 share the torch~=2.13.0 pin, and torch 2.13.0 still publishes cu126 wheels, so Volta keeps a supported route; 2.15 drops cu126 entirely. - FRAMEWORK_PYTHON gains torch 2.13.0 and _torch_modern_cuda maps it to cu130, the CUDA version of the default PyPI torch 2.13.0 build (cuda-toolkit 13.0.3). CUDA 13 carries no kernel below Turing, which matches the modern-architecture set; an unknown torch version still fails closed. - UMACalculator now owns the preset mapping: mliport "default" runs upstream "batch" (tf32 off, no compile, no merged MOLE = the 2.21 "default" numbers) and "turbo" runs upstream "multigpu" (upstream calls it "turbo's former settings"). A preset that would enable Nitro is refused, a missing preset fails closed with an upgrade message instead of degrading, and the effective preset, its tf32 flag and its Nitro flag are added to the model identity. - Generated install docs and the UMA backend page are regenerated; the page records that the beta scientific validation of uma_omat ran on 2.21.0 and does not carry over. - Tests updated for the new pins and new tests pin the preset mapping, the Nitro refusal and the fail-closed missing-preset path. Evidence (owner's V100, Tesla V100-SXM2-16GB, driver 580.173.02): ./scripts/install_mliport.sh --engines uma --device auto re-resolved the existing .venv and its closing `mliport doctor --engine uma --device auto` passed: fairchem-core 2.25.0, torch 2.13.0+cu126 (CUDA 12.6), a real tensor dot product on cuda:0, arch list sm_50/sm_60/sm_70/sm_75/sm_80/sm_86/sm_90. The UMA checkpoints themselves are not called here: facebook/UMA is a gated repository and this host has no HF token, so uma-s-1p2p1*.pt could not be downloaded. That callable check remains not_run. DATA IMPACT: no stored scientific record invalidated. Runs produced on fairchem-core 2.21.0 keep their recorded framework versions and stay valid for that runtime; they are not interchangeable with runs on 2.25.0, which the docs now say explicitly.
# Conflicts: # CHANGELOG.md
Tracking: #61 (model-support spec §11 task 4, review of #97) What was missing: the docs said Turing and newer use "cu128 or later" and that the UMA runtime is on cu130, but not that the CUDA 13 wheels require an NVIDIA driver of 580 or newer, and not that Volta stays on cu126 precisely so older drivers keep working. What changes: - `docs/installation.md` (+ `.zh.md`): the GPU-architecture section names the `cu130` driver floor and says a CUDA 13 build is not an alternative for pre-Turing GPUs because it has no kernels below Turing, whatever the driver. - `docs/backends/uma.md` (+ `.zh.md`): the "Runtime pin" bullet says the same in one sentence. - CHANGELOG entry. Verified on the owner's host: driver 580.173.02 runs torch 2.13.0+cu126 on the V100 (sm_70 kernels present in the wheel's arch list). DATA IMPACT: no scientific record invalidated; documentation only.
|
§11 task 4 done, and task 1 has passed for the checkpoints that are readable on this host, so this PR is marked ready for review. Changes requested by §11 task 4:
Task-1 boundary, for the reviewer: on the #97 runtime |
Resolve the CHANGELOG `Unreleased` conflicts by keeping every entry (registry corrections, MACE gate, and this branch's UMA runtime note) and dropping a whitespace-only hunk. No code file conflicted. DATA IMPACT: none. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GYQHUBfRm75zyaocAPWhEb
Tracking: #61 (model-support spec §6.5 + §11 tasks 1 and 5) Rebased onto `main` after #94, #96 and #97 merged, so this carries only the status-table change. What changes: - The status table leaves its initial state and carries what was measured on 2026-11-14 on the owner's V100 in the installer-built environments: 18 profiles `verified` on CPU and GPU, each with finite energy/force/stress on the 50-atom LGPS cell, a matching recorded identity and the bounded probe (shared calculator, central differences, 10 NVE steps, five-image NEB). - UMA rows from §11 task 1: `uma_omat`, `uma_s_1p2p1_omat` and `uma_s_1p2p1_ft_r2scan_omat` verified on both devices and recorded with `inference_preset: batch`, `nitro_enabled: false`, `tf32_enabled: false`; `--inference-mode turbo` records `multigpu` with `nitro_enabled: false`. The r2SCAN fine-tune's NEB leg is refused by the gradient preflight (energy resolution), which the table states instead of hiding. `uma_omat` needed the re-uploaded checkpoint: the first copy had the right size but failed the zip CRC; the new file matches the pinned manifest hash `ba5c0d912efa...`. - The CPU-GPU claim quotes the measured maxima (6.1e-7 eV/atom, 7.5e-6 eV/Angstrom, `mace_omat_float32`) instead of a rounded "1e-7/1e-5 or better". - Every Evidence cell links the specific #75 comment (runs, deltas, UMA record, re-check, NEP89 record) instead of the pull request. - The Chinese table is translated column by column; it previously carried the English dtype and evidence cells. DATA IMPACT: no scientific record invalidated. Statuses describe calling the models correctly; no accuracy claim is made and no stored campaign record changed.
Tracking: #61 (model-support spec §6.4). Draft: ready for independent review.
What was wrong
The pinned fairchem-core 2.21.0 cannot load the UMA-S 1.2.1 checkpoints. The bump is not
only a version change: 2.22+ made UMA Nitro the upstream
defaultpreset and removedturbo, so re-pointing the pin would have silently changed the precision of everymliport "default" run and could have selected Nitro (Ampere+, reduced precision).
What changes
fairchem-core==2.25.0ontorch 2.13.0(torch~=2.13.0, Python >=3.11,<3.15).Volta keeps
cu126; Turing+ usescu130, the CUDA version of the default PyPItorch 2.13.0 build (
cuda-toolkit==13.0.3); CUDA 13 has no kernels below Turing.FRAMEWORK_PYTHONgains torch 2.13.0;_torch_modern_cudamaps it tocu130. Anunknown version still fails closed.
default→ upstreambatch(tf32 off, no compile,no merged MOLE = the 2.21
defaultnumbers),turbo→ upstreammultigpu(upstream's "turbo's former settings"). A preset that sets
nitrois refused; amissing preset fails closed with an upgrade message.
identity. UMA Nitro is not exposed.
docs/backends/uma.md(+.zh.md) states that the beta scientificvalidation of
uma_omatran on 2.21.0 and does not carry over.Evidence (owner's V100)
./scripts/install_mliport.sh --engines uma --device autore-resolved.venvand itsclosing
mliport doctor --engine uma --device autopassed: fairchem-core 2.25.0,torch 2.13.0+cu126, real tensor product on
cuda:0, arch listsm_50/60/70/75/80/86/90. The UMA checkpoints themselves were not called: facebook/UMA
is gated and this host has no token (
401), souma-s-1p2p1*.ptisnot_run, not passed.Tests
Updated plan/doctor/gpu-setup/architecture tests for the new pins; new tests pin the
preset mapping, the Nitro refusal and the fail-closed missing-preset path.
DATA IMPACT
No stored scientific record invalidated. Runs produced on fairchem-core 2.21.0 keep their
recorded framework versions and remain valid for that runtime; they are not
interchangeable with runs on 2.25.0, which the docs now say.