requirements: record torch 2.8.0 / 2.11.0 as checked against 2.6.0 - #428
Open
siyuanj wants to merge 1 commit into
Open
requirements: record torch 2.8.0 / 2.11.0 as checked against 2.6.0#428siyuanj wants to merge 1 commit into
siyuanj wants to merge 1 commit into
Conversation
Both were verified against torch 2.6.0 on sm_89 (L4) and sm_100 (B200) with chai_lab 0.6.1: the traced ESM2-3B forward pass is bit-identical, and a full deterministic-mode fold of 1URN reproduces the same coordinates and the same ptm / iptm / aggregate_score. Details in chai-lab#427.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Comment-only change. Follows up on #427, where the practical problem is that the released
0.6.1wheel still pinstorch<2.7and therefore cannot be used on any Blackwell GPU.mainalready dropped that upper bound in #383, with the comment recording 2.3.1 – 2.7.1 as confirmed. Since running past a documented range is uncomfortable when you can't tell whether it perturbs numbers, I measured it, and this records the result next to the requirement so the next person doesn't have to repeat the work.What was checked, with
chai_lab0.6.1 and only the torch wheel swapped (apython -m venv --system-site-packageslayer over one shared install, sochai_lab/numpy/rdkit/einops are literally the same files — thepip listdiff is torch + triton +nvidia-*-cu12+ sympy and nothing else):6f1c44c9e3b6768bf9167fe79bb92381d45b0a5fa3acf1c918f46ca7ef6b4a9c, 0 of 248 320 elements differingchai-lab foldon 1URN underuse_deterministic_algorithms(True)+CUBLAS_WORKSPACE_CONFIG=:4096:8,--seed 42, same L4: torch 2.6.0 and 2.8.0 give byte-identical coordinates for all 5 × 1225 atoms; 2.11.0 gives bit-identicalptm/iptm/aggregate_scoreand coordinates differing only in the last written CIF decimal on 23–60 of 1225 atoms (max 0.0014 Å)sm_89) and B200 (sm_100), driver 580.159.04I deliberately did not widen the existing "2.3.1 - 2.7.1 are confirmed to work correctly" range to
- 2.11.0: I only tested 2.8.0 and 2.11.0, not 2.9.x or 2.10.x, and only on one protein–RNA target without ligands, templates or constraints. So this adds a separate, narrower claim rather than extending yours to versions neither of us has run.Happy to reword or drop it if you'd rather keep the requirement line uncluttered — the data in #427 is the part that matters.