commits on main
← newer commits2026-10-17
f15273eb dawn Round 5 blitz: gr_norm_h_moec/gr_apply_moec, qsa_prep#qf16, -580 MB scratch aliasing, deterministic prompt lookahead (352.1 tok/s prefill, 1.58 GB scratch)
Round 5 blitz: gr_norm_h_moec/gr_apply_moec, qsa_prep#qf16, -580 MB scratch aliasing, deterministic prompt lookahead (352.1 tok/s prefill, 1.58 GB scratch)
89188acf dawn Round 4 blitz: gr_norm_h_moe + PLE fusion (-189 MB scratch), mtp_rms_xprep + mtp_combine_gnorm (218 ms MTP), 5.9x MoE plan + 1.7x mm_gather_h (336.4 tok/s prefill, 32.8-41.0 tok/s decode)
Round 4 blitz: gr_norm_h_moe + PLE fusion (-189 MB scratch), mtp_rms_xprep + mtp_combine_gnorm (218 ms MTP), 5.9x MoE plan + 1.7x mm_gather_h (336.4 tok/s prefill, 32.8-41.0 tok/s decode)
b3decf33 dawn Round 3 blitz: GDN decode/prefill kernel optimization (gdn_step -12%, gdn_scanr_4 -8.4%), r1 MoCE unscaling fix, spec_min_p schedule (32.3-40.3 tok/s decode, 325 tok/s prefill)
Round 3 blitz: GDN decode/prefill kernel optimization (gdn_step -12%, gdn_scanr_4 -8.4%), r1 MoCE unscaling fix, spec_min_p schedule (32.3-40.3 tok/s decode, 325 tok/s prefill)
4b0f233f dawn Round 2 blitz: gemv_cat+gr_norm_comb fusion, 1247 prefill barriers removed, M5 gemm_f16 tiles, slop4 MTP, P_GSLOT cache fix, parallel qsa_prep (32.4-41.2 tok/s)
Round 2 blitz: gemv_cat+gr_norm_comb fusion, 1247 prefill barriers removed, M5 gemm_f16 tiles, slop4 MTP, P_GSLOT cache fix, parallel qsa_prep (32.4-41.2 tok/s)
bec6ab3b dawn Round 1 blitz: MTP batched fc_hidden + sink=64 + valid-run KV, single-rw mm_gu3, multi-tok M5 gemv/moe tiles, prompt publish_ready (32.7-40.8 tok/s)
Round 1 blitz: MTP batched fc_hidden + sink=64 + valid-run KV, single-rw mm_gu3, multi-tok M5 gemv/moe tiles, prompt publish_ready (32.7-40.8 tok/s)
d6a726f8 dawn Protect 75% of Class 0/1 arena slots during warm multi-turn prefill deltas (pos0>0)
Protect 75% of Class 0/1 arena slots during warm multi-turn prefill deltas (pos0>0)
7f8dc234 dawn Speed up warm multi-turn prefill deltas (GEMM_MIN=128 + pos0>0 Tail-16/Prot-3 PROMPT_MOCE)
Speed up warm multi-turn prefill deltas (GEMM_MIN=128 + pos0>0 Tail-16/Prot-3 PROMPT_MOCE)
f6e2147b dawn Add Slim MTP + AcceptMoE draft routing (30.4-42.5 tok/s at 32K context, 6.5 GB arena, 256-cell MTP window)
Add Slim MTP + AcceptMoE draft routing (30.4-42.5 tok/s at 32K context, 6.5 GB arena, 256-cell MTP window)
a7a5f5b0 dawn Restore 6.8 GB wired arena at 32K context (CLM sweet spot), PROMPT_GEMM_MIN=64, and bounded 2-slot SCR ck_ring
Restore 6.8 GB wired arena at 32K context (CLM sweet spot), PROMPT_GEMM_MIN=64, and bounded 2-slot SCR ck_ring
0d1cab19 dawn Add native 49th-layer neural MTP (27.8-29.4 tok/s across 4 domains), QSA tail_bak rollback, and stripped-think SCR in --serve
Add native 49th-layer neural MTP (27.8-29.4 tok/s across 4 domains), QSA tail_bak rollback, and stripped-think SCR in --serve
411096fc dawn Add prefill lookahead router + async PLE (335-346 tok/s @ 2K), 64K/128K decode benchmarks (20.4-20.8 tok/s), and QSA-safe suffix drafting
Add prefill lookahead router + async PLE (335-346 tok/s @ 2K), 64K/128K decode benchmarks (20.4-20.8 tok/s), and QSA-safe suffix drafting
9d4fcca5 dawn Add Anti-Slop v4.1b, M5 decode/prefill optimizations (22.64 tok/s decode, 5.27x prefill), and SCR pad-carry
Add Anti-Slop v4.1b, M5 decode/prefill optimizations (22.64 tok/s decode, 5.27x prefill), and SCR pad-carry
2026-10-16
8e88dda1 dawn Add unpruned K=512 ukisai GSQ-RCO IQ2_XS Metal builder and live server script (21.00 tok/s)
Add unpruned K=512 ukisai GSQ-RCO IQ2_XS Metal builder and live server script (21.00 tok/s)
ad81eaff dawn Add --serve mode to strata_window_cli with prefix/SCR reuse and Albedo local provider wiring
Add --serve mode to strata_window_cli with prefix/SCR reuse and Albedo local provider wiring
977afaa3 dawn Complete Phase 5 Disk Session + SCR 36-layer GDN staleness & 190-pair Real-VN anti-overfitting audit
Complete Phase 5 Disk Session + SCR 36-layer GDN staleness & 190-pair Real-VN anti-overfitting audit
be8f7b6f dawn Complete Phase 4 v3.2 Layer-47 WFLD anti-slop, Phase 5 Strata Metal v3 (22.08-22.25 tok/s), and Opus audits (v1-v3.2 + blind sign-off)
Complete Phase 4 v3.2 Layer-47 WFLD anti-slop, Phase 5 Strata Metal v3 (22.08-22.25 tok/s), and Opus audits (v1-v3.2 + blind sign-off)
2026-10-15
9220caa0 dawn Complete Phase 3 local M5 healing, GGUF transplant, and 22-window/9-prompt generalization audits (v1-v3 PASS)
Complete Phase 3 local M5 healing, GGUF transplant, and 22-window/9-prompt generalization audits (v1-v3 PASS)
4d17890a dawn Save Phase 0–2 M5 validation, Phase 1a–1c REAP/MoCE results, Phase 3–4 prep scripts, and Opus audits
Save Phase 0–2 M5 validation, Phase 1a–1c REAP/MoCE results, Phase 3–4 prep scripts, and Opus audits
2026-10-14
full history: git clone https://git.klbr.net/ptr.pet/regolith