Buckets:
| """Mamba-3: a paper-faithful, verified implementation of Lahoti et al. (arXiv 2603.15569). | |
| Ported from the MMSE research repo (drowsiness-uta-rldd/mamba3, commit c5280e6+), without the MMSE | |
| multi-stream extension or the LM tooling. Checked fwd+bwd at float64 against the recurrence: | |
| - exponential-trapezoidal discretization with DATA-DEPENDENT lambda_t (Prop 1) | |
| - complex state as real block-diagonal rotations / data-dependent RoPE (Prop 2-4, eq 11) | |
| - MIMO rank-R (Sec 3.3) | |
| - BCNorm/QKNorm on B,C + learnable data-independent B,C biases | |
| - output gating | |
| Recorded deviations: data-independent A (Remark 1 says equivalent); MIMO input uses a full DPR | |
| projection rather than the paper's DP+PR (R=1 identical). | |
| Paths: `Mamba3Layer.forward` (O(T) recurrence, reference), `scan_forward` (chunked scan), | |
| `scan_forward_ssd` (SSD reduction, the fast path for training). Run `python -m | |
| steertts.mamba3.tests` after any change. | |
| """ | |
| from .layer import Mamba3Layer, rotate_pairs, ssm_scan | |
| from .scan import scan_forward, chunked_scan, rotate_by, cumulative_angles | |
| from .ssd import scan_forward_ssd, ssd_inputs, ssd_pytorch | |
| __all__ = [ | |
| "Mamba3Layer", "rotate_pairs", "ssm_scan", | |
| "scan_forward", "chunked_scan", "rotate_by", "cumulative_angles", | |
| "scan_forward_ssd", "ssd_inputs", "ssd_pytorch", | |
| ] | |
Xet Storage Details
- Size:
- 1.31 kB
- Xet hash:
- 6de8be4a18bac88630c3663fb76e1b6370d4ac6922d0faa05b72a57f1430a326
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.