Residual Transport
Training campaign launched 2026-10-10. Training is ongoing; no completed RT results are claimed.
Source: https://github.com/anhdao69/LoopWAM_NT/tree/residual_trans
Implementation revision: f669184242b56f24ff07fa2c987c799036723a0b
- RT-A: batch job 4847; resumes saved update 220.
- RT-B4: batch job 4846.
- RT-T1ft: batch job 4849.
- RT-Pa: batch job 4850.
- RT-TF: batch job 4851.
- All seven runs are independent, with no dependencies or evaluation waits, per the user's later instruction.
- RT-A2: standalone one-H100 batch job 4861, resuming update 58; source a7403b5aab91, microbatch 64 with accumulation 2.
- RT-B2a: two-H100 batch job 4859; source a7403b5aab91, microbatch 64 with accumulation 1.
- These two phase-2 launches also override their evaluation wait at the user's request; no Gate 1 result is claimed.
- Remaining 13 executable recipes retain their gates. R-FM remains undefined and deferred.
- All seven runs now use standalone batch jobs. Interactive allocation 4857 was canceled after saving RT-A2 at update 58. There is no additional backup job.
Every run uses global batch 128, all 277,713 windows from 1,712 available LIBERO demonstrations, and ten epochs (21,700 updates). RT-A2 uses one H100; the other six submitted runs use two H100s each. Automatic retrying uploads publish each run's update-2,000 early checkpoint and separate epoch_08.pt, epoch_09.pt, epoch_10.pt files with configuration and dataset metadata. Recovery optimizer/RNG files and synthetic smoke weights are not published.
Checkpoint use
RT weights require the TransportInference adapter from the source repository. Use --transport-schedule S1 or S2 with scripts/evaluate_loopwam_libero.py for phase-1 evaluation. The exact unmerged trained action state is retained alongside a merged backbone to preserve BF16 inference rounding. Use the matching dataset_stats.json and Wan2.1 VAE listed in config.json. Checkpoint absence means the corresponding training milestone has not yet been reached.
Verification and gates
328 project tests passed, seven skipped. All 20 recipes passed native GPU forward/backward smoke checks. Ten resumed updates matched uninterrupted training bit for bit. 26 graph cases matched eager inference across 50 changing queries each. Export/reload and B2a initialization checks passed. Two/four logical-rank partition parity was tested on two physical GPUs; four-device NCCL was not tested. No full simulator evaluation ran on the H100 training server. The experiment matrix and launch report describe the preserved evaluation gates.
This is a dated launch snapshot. Completed checkpoint files and per-run metadata are the authoritative publication state.
Single-GPU support passed 30 focused tests and a 20-update global128 smoke across all six RT-A2 schedules. Ten resumed single-GPU updates matched uninterrupted weights, Adam, EMA, losses and gradient norms bit for bit.