Eval mode 2 does part-wise Rel-L2 averaging, not joint full-mesh Rel-L2

[vpuri3]

Observation

Eval mode 2 (--eval 2: decoupled inference / physical state caching + full mesh decoding) does not appear to do joint prediction + Rel-L2 over all cells in a mesh. It does piecewise predictions over parts, then averages per-part Rel-L2s.

That conflicts with Appendix A.3 Eq. 7 (Rel-L2 over the full mesh of (N) cells).

Evidence

Test-mode dataloader yields one part at a time, not the entire mesh:

https://github.com/thuml/Transolver-3/blob/main/dataset/dataset_drivaerml_surface_numpy_chunk.py

• train=False: length is #runs × #parts
• __getitem__ returns a single stride-part

train_surface.test and test_decoupled_inference Stage 2 then compute Rel-L2 on that part and np.mean over batches:

https://github.com/thuml/Transolver-3/blob/main/train_surface.py

So reported L2RE is mean_parts(RelL2(part)), not joint RelL2(entire mesh).

Mode 2 Stage 1 builds a global cache, but Stage 2 still decodes/evaluates part-by-part with that averaging.

Question

Were published surface Rel-L2 numbers computed this way, or via true full-mesh accumulation? If Eq. 7 is intended, eval should sum error/target energies across parts (per run) and take one ratio.