LUSA and companion selector weights
LUSA: Lightweight Utility-Guided Sparse Attention for Long-Context Prefill
Paper1 final selected weights
Seven selected settings: two backbone models and five ablations. The original files listed below use the historical phase_tiny_selector/v3 format and require the matching historical runtime. The current public LUSA runtime 7ed144b accepts only lusa_selector/v1; use the corresponding files under paper1/lusa_v1/<variant>/selector.pt. The two formats preserve the same selector tensors.
paper1/minicpm_kl21/best-step140.ptpaper1/minicpm_fixed_mixing_s42/selected-best-step20.ptpaper1/minicpm_listwise_only_s42/selected-best-step260.ptpaper1/minicpm_no_boundary_s42/selected-best-step60.ptpaper1/minicpm_no_worst_s42/selected-best-step180.ptpaper1/minicpm_single_subspace_s42/selected-best-step180.ptpaper1/nosa_kl21/selected-best.pt
Load selector PT files with torch.load(path, map_location="cpu", weights_only=True) and use the matching sparse-model integration. These selector files require the original frozen backbone.
Inventory and companion collections
The root manifest contains 72 weight paths across all collections. Provenance maps original and published checksums.
attention/sparda/baseline-step2000/indexer-deploy.pt contains projection tensors and inference-only sparse_config; load model_state_dict into the matching MiniCPM sparse implementation.
attention/depth-attention/baseline-step50000/model.safetensors requires the matching custom Depth-Attention implementation and a compatible tokenizer. Its accompanying configuration is provided.
Apply each companion artifact's upstream license. Paper1 results, analyses and manuscripts are stored in the private research repository. Selected Paper1 files exclude optimizer/RNG state. Companion collections also contain historical pilot and intermediate selectors; see the inventory before using them.
Inference-only cleanup (2026-10-30)
Thirty-seven companion checkpoints were converted in place to inference-only files. Model tensors and sparse configurations were verified against the previous public files; optimizer, scheduler and RNG state were removed and training resume is not supported. Their filenames are unchanged but their SHA-256 hashes changed. The seven selected Paper1 deployment weights were not altered. See conversion receipt and the updated manifests. This verification does not claim a new GPU inference or benchmark run.
The additional conversion receipt records 19 compact QAMS, HCR and MCI files whose optimizer and scheduler keys were also removed.
Historical CPT P2 deployment weights
cpt-p2-128k/ retains the CPT P2 iter-300 deployment export identified by the September 2026 archive as its latest deployment checkpoint. The two model shards total 17.31 GB. Their LFS hashes are unchanged; a training-machine path in the configuration was redacted. The historical static check found 12,513 indexed tensors and identical tied embeddings; the index total-size metadata differs by one tied embedding. GPU loading and inference were not tested for this migration. See migration receipt. Download this directory explicitly when needed, e.g. hf download Xnhyacinth/sparda-checkpoints --include "cpt-p2-128k/*".
Historical branches and recovery
Historical branch heads were archived as shallow Git trees in the private research archive. That archive contains Git files and LFS pointers, not LFS payloads. Selected research results, checkpoint mappings and intentional historical LFS removals are documented there; old historical branch URLs and checkpoint bytes may no longer be available.
Which checkpoints are final?
The historical 95-file .pt cohort included seven selected Paper1 selector modules, companion weights, early ablation pilots and historical experiment checkpoints; they are not 95 complete final models. The selected modules require a matching frozen backbone and implementation. Other historical settings have not all been validated for current inference. See the checkpoint inventory for the full classification and exact hashes.
Canonical LUSA compatibility
The conversion map links each historical selected selector to its lusa_selector/v1 file and exact source/output SHA-256. All 128 tensors in each model and the numerical configuration were compared. The current LUSA loader and CPU mock-layer installation passed for all seven; the recorded MiniCPM/NOSA backbone configs and NOSA routing/backbone manifest were checked. This is not a new full-model or GPU inference evaluation. The original seven v3 files remain for historical integrations.
For paper-used settings, historical Tiny and SparDA baselines, and the distinction between Uniform and learned Mix, use the paper and baseline checkpoint map.
The current tree has 69 .pt paths, including seven canonical LUSA encodings and the six historical 50-step ablation endpoints. Two smoke weights and 18 initialization/intermediate paths were retired. The full_matched step-zero was used for a historical untrained control, but its weight has been retired under the final-model-only policy; its result and identity record remain.
Six additional nonfinal training snapshots were retired; their historical identities and provenance remain recorded.
Four step_002000.pt endpoint containers were retired after recursive tensor/config equivalence with their retained final.pt; differing metric fields are preserved in the private research record.
Two qualification-only temporal checkpoints and one failed temporal-v1 checkpoint were retired; the distinct completed setting and historical result notes remain.