Skip to content

orbmol configs: name the arch for the :custom smoke-test leg - #231

Open
OwenPriceSkelly wants to merge 1 commit into
mainfrom
orbmol-custom-leg-arch
Open

OwenPriceSkelly wants to merge 1 commit into
mainfrom
orbmol-custom-leg-arch

Conversation

@OwenPriceSkelly

Copy link
Copy Markdown
Member

Problem

The smoke-test's weights= leg (#200, #204) borrows the first fetched same-family checkpoint in CHECKPOINTS order and reloads its weights through setup_from_path. For orbmol that baseline is orbmol-v1-conservative, but setup_from_path defaults to arch="orbmol-v2". The v2 energy head has 257 input columns (the CoulombModule adds one) against v1's 256, so the leg fails:

RuntimeError: Error(s) in loading state_dict for ConservativeForcefieldRegressor:
    size mismatch for heads.energy.mlp.NN-0.weight: copying a param with shape torch.Size([256, 256]) from checkpoint, the shape in current model is torch.Size([256, 257]).

Seen on Frontier's 2026-08-24 run (orbmol:custom red on the Almanac since). Both the amd_configs and nvidia_configs copies have the same default.

Fix

Add a VERIFY_KWARGS entry keyed by the :custom id, which the leg already falls back to (setup_kwargs={} → verify_kwargs_for), naming the loader that matches the borrowed weights. Users passing their own weights still choose arch= per call; nothing else changes.

Verification

On Frontier login03, 2026-09-29, after appending the same block to the deployed env files:

orbmol/orbmol-v1-conservative   [PASS]  cuda  105.0s
orbmol/orbmol:custom            [PASS]  cuda   50.6s  (weights from orbmol-v1-conservative)

tests/sample_configs and tests/verify/test_verify_kwargs.py: 166 passed.

🤖 Generated with Claude Code

The weights= leg borrows the first fetched family checkpoint (declaration
order: orbmol-v1-conservative) and reloads its weights through
setup_from_path, whose default arch is orbmol-v2. The v2 energy head has 257
input columns against v1's 256, so the leg failed on Frontier with
'size mismatch for heads.energy.mlp.NN-0.weight' (2026-08-24 run). A
VERIFY_KWARGS entry keyed by the :custom id selects the matching loader for
the leg only; users still pass arch= for their own weights.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant