3.0 KiB
3.0 KiB
Transolver-Greif Adapter (TRANSOLVER-ADAPTER-1122)
Code-only Transolver (Wu et al., ICML 2024) adapter aligned with the frozen grouped Greif v5 comparison protocol (18e0e729). No real data read, no GPU run, no push to main.
Files
tools/track_b_transolver_greif_adapter.py:GreifTransolver,TransolverBlock,PhysicsAttention,make_greif_transolver, and pure-array input helpers.tools/test_track_b_transolver_greif.py: CPU-only synthetic tet-mesh unit tests.
Protocol Alignment
- Inputs: Exact 19
COLUMNSfromtrack_b_gino_greif_adapter.py(no extraxyzchannels), standardized in FP64 on each fold's learning split only (phase2_h2_data.standardization), cast to FP32, plus strictly positive FP32 nodalmassand optionalcomponent_ids. - Target & Decode: Exact
phase2_residual_mathFP64 volume-clock priorc * Vf, signed unclamped FP32residual_target, anddecode_seconds. Readout head final layer is zero-initialized so epoch 0 starts atc * Vf. - Loss & Schedule: Reuses
phase2_h2_model.{weighted_loss, optimizer, lr_epoch}(node-equal FP32 MSE, AdamWwd=1e-4on >=2D non-bias, 10-epoch warmup + 190-epoch cosine, grad clip 1.0, TF32 off). - Folds: Unchanged grouped folds 0/1 (
((0,0),(0,1),(1,0),(1,1)), family isolation,no_validation_selection=True).
Parameter Counts
| Model | Config | Parameters |
|---|---|---|
H2Geo (reference) |
C=64, 4 GeoBlocks |
195,077 |
GreifTransolver (default) |
C=64, L=6, H=4, G=64, mlp_ratio=2 |
197,913 |
GreifTransolver (wide flag) |
C=96, L=4, H=4, G=64, mlp_ratio=2 |
293,937 |
Design Assumptions
- Mass-weighted slice pooling (
mass_weighted=True, default): Nodal mass is normalized per connected component (m / sum(m)) before slice pooling so attention tokens are invariant to global mass units and reflect tet volume measure.mass_weighted=Falsereproduces unweighted Transolver. - Per-component isolation & chunking: Slices are pooled per physical connected component (matching
H2Geo.global_message) inNODE_CHUNK=8192chunks with optional block gradient checkpointing to bound VRAM on 30K–100K+ node tet meshes. - Slice denominator & temperature stability: Per-head temperature is clamped to
[0.1, 5.0]and normalized slice denominators add1e-6(1e-5unweighted) to prevent FP32 division instability.
Open Items for MeltSol / MeltAssist
- Execution binding & worker: This branch provides only the model adapter and CPU synthetic tests; MeltSol should wire
GreifTransolverinto a guarded fold-0/1 worker/entry (samefirst-epoch cost gate and <= 40,000 GPU-s GPU3 cap as planned) for MeltAssist review. - Checkpoint / VRAM profiling on real meshes: Verify
NODE_CHUNK=8192andcheckpoint_blocks=Truepeak VRAM on the largest Greif learning part during the first measured epoch. - Optional ablation: Confirm whether only
default(197.9K,mass_weighted=True) runs in the primary <= 40,000 GPU-s slot, leavingwideandmass_weighted=Falseout of scope.