v1.3.2

What's Changed

Repository Stats

Component	v1.3.1	v1.3.2	Added
Skills	127	127	—
Workflows	9	9	—
MCP Tools	49	49	—
Tool Servers	20	20	—

Other Highlights

• NValchemi batched MD gating: Re-benchmarked sequential vs. batched MD serially (no GPU contention) and found NValchemi batched MD is a win only for MACE (4.90x). Added a _nvalchemi_supports_batch_md flag that routes TensorNet and FairChem MD to the faster sequential path. TensorNet's light forward exposes a torch.compile/Warp neighbor-list stream race (officially supported by matgl.ext.alchmtk, so a hardware limitation, not an incompatibility) and is ~0.88x even when forced; FairChem uma-s-1p2's forward scales superlinearly per atom (1.57 ms/atom at batch=1 → 2.43 at batch=20) so batched MD is 0.64x.
• FairChem cutoff fix: Corrected the NValchemi FairChem wrapper's cutoff inference, which fell back to 12 Å / 500 neighbors instead of the model's true 6 Å / 300 (wrong attribute paths), making adapt_input build ~530 edges/atom of work that uma-s-1p2 discards (it runs with external_graph_gen=False). Added a robust _find_backbone walk.
• NValchemi fixed-batch relaxation: Switched to standard form cell mapping and scaled FIRE2 optimizer to prevent cell distortion during relaxation. Fixed trailing post-convergence step handling and enabled native convergence masking in fixed-batch mode. Disabled inflight batching for TensorNet to prevent neighbor-list OOB errors.
• MCP Server memory optimization: Implemented automatic GPU memory cleanup (gc.collect() and torch.cuda.empty_cache()) within MACE, MatGL, and FairChem relaxation and molecular dynamics handlers to prevent VRAM accumulation.
• Skills & documentation: Re-documented the NValchemi skill with corrected batched vs. sequential MD benchmarks for MACE, MatGL (TensorNet), and FairChem (UMA), and added a serial run_md_benchmark.py harness. Refined category tags on mass spectrometry and spectrum-matching skills.