audits · regolith · Tangled

strata but optimized for m5 (and maybe more later)

regolith audits

2 folders · 135 files

dawn · audits: r45 colab ssh-slot refusal - ColabSshSessionError + HTTP recovery helpers (patch in dawn's albedo checkout, uncommitted) · 1w ago

r44_chunk0_moce

audits: archive r44 chunk-0 MoCE measurement (reconstructed; 47/64 KL rows unrecoverable, hashes null)

r45_colab_ssh_slot

audits: r45 colab ssh-slot refusal - ColabSshSessionError + HTTP recovery helpers (patch in dawn's albedo checkout, uncommitted)

argon_k512_decode_opt.md

Add Anti-Slop v4.1b, M5 decode/prefill optimizations (22.64 tok/s decode, 5.27x prefill), and SCR pad-carry

argon_multi_tok_kernels.md

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)

argon_prefill_barriers.md

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)

argon_prefill_plan.md

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)

argon_r10_gather.md

Round 10 blitz: qsa_flash 2.2x (-51.9% / -146.1 ms in kprof via unconditional hoisted rv/rk + vec4 epilogue), gdn_pre float4 conv, uint4 mm_gather_h + s_id-free mm_place

argon_r11_moemm.md

Round 11 blitz: gemm_f16 tile selector (-18% / -234 ms), qsa_attn_mm hoisted rv/rk (-44.5%), branchless W8<T_IQ3_S> (-50.8% / 2.03x) + W8<T_IQ2_XS> k_hoist (-57 ms), 3,397.5 ms GPU kprof (583.7 tok/s GPU) & 2048 prompt chunk on :8080

argon_r12_gemm.md

Round 12 blitz: qsa_topk 3.85x (cached keys + simdgroup scan + paired emit), gemm_hc_up_reg_32_128_4 + unrolled epilogue (-6.9%), Q6_K hoisted mm_gu3 (-4.3%), 3,375.4 ms GPU kprof (587.5 tok/s GPU)

argon_r13_dqflat.md

Round 13 blitz: gemm_hc_down + gemm_swiglu_up K4 register epilogues (-144 dispatches, -985 MB DRAM/chunk, 600.6 tok/s GPU kprof), dq_f16 exact-simdgroup 2D grid, SpecGate + mtp_draft PLE prefetch

argon_r14_gdn_gr.md

Round 14 blitz: qsa_flash/qsa_attn_mm byte-offset + zero-scale hoist (-9.7% at pos0>=2048, 443.2 tok/s on 10.18K Albedo prompt), gdn_scanr_4 hk-locality, compacted moe_comb roff, mm_gu3 IQ2_XS half4 HS store

argon_r15_prep.md

Round 15 blitz: prompt_chunk router+plan/shexp wave zipping (-190 barriers), fused Wal+Wbe N=96 GEMM (-36 dispatches), QSA PSO pre-warm (-202 ms Chunk-1 stall), qsa_score_mm causal-future skip (-19.4%), qsa_prep 320B TG + 2 barriers

argon_r16_ple.md

Round 16 blitz: fused QSA Wk+Wv+Wik(+Wiq) single GEMM (-24..-36 dispatches, -5.6 MB scratch), decode resident-MoE + shared-expert wave zip (-48 barriers), gemv TAB>1 guards, ple_moe_gate_norm_c SIMD compaction + ple_conv half4/t>=9

argon_r17_gdn.md

Round 17 blitz: mtp.metal float4 vectorization + Q8_0 gemv_catk, barrier-free simd_shuffle_down in gr_up_f/gr_up3/qsa_merge, gdn_step ib==0 readout pruning + qsa_attn char4 K loads (39.56 tok/s warm decode)

argon_r18_qsa2.md

Round 18 blitz: per-stream gr_norm_comb 5.09x (-80%), xprep float4 (-46.8%), moe_combine_sh (-49%), Q6_K moe_gu NE>=3 (-25%), qsa_topk2/qsa_score (-5..9%) -> 40.03 tok/s decode & 445.79 tok/s 10.18K prefill

argon_r19_gemv.md

Round 19 blitz: gemv NTOK/NR/KS unroll (-5.4% to -29.4% at ntok>=2), mtp_pick/slop unroll_count(4) (-27..60%), ple_comb_query_gate_norm + load-time prewarm -> 40.44 tok/s decode & 212.5 ms mtp_ms

argon_r20_gdn.md

Round 20 blitz: mtp_load mlock 708 MB experts (mtp_ms 398->122 ms on 11K, 433->202 ms on 2K; 40.57 tok/s decode), qsa_attn (-20..40%) + qsa_merge (-10..13%), gdn_scanr_4 32-bit offsets (186.15 ms kprof)

argon_r21_gdn_moe.md

Round 21 blitz: gr_down/gr_up GR_NTOK unroll (-5.4..18.0%), qsa_attn2/qsa_merge2 (-36.8..40.6% / -8.3..14.6%), gdn_step st_off + moe_plan topk10 unroll_count(10) (-9.5..26.6%) -> 41.66 tok/s decode & 193.9 ms mtp_ms

argon_r22_moe.md

Round 22 blitz: gr_down_4_4/_r2 NR=2 xn reuse (-8.1..26.1%), moe_down/moe_gu q2_combine (-7.7..22.5%) + plan_body simd_prefix (-14.6..35.4%), swiglu_prep barrier-free -> 35.66 tok/s coding & 24.71 tok/s @10.2K decode

argon_r23_moemm.md

Round 23 blitz: ksg popcount + gemv_direct_x/gemv_q2_raw_combine + mm_gu3 sign4_iq2xs/ptr-folding -> 42.01 tok/s decode & 602.1 tok/s GPU prefill

argon_r24_moemm.md

Round 24 blitz: W8<T_Q6_K> (+12.8..24.9%), qsa_flash 32-bit ko/vo (-11.7% @8K), TAB 1024/512 + gr_norm_h_moec k-outer + gemm_hc_down *0.25f

argon_r25_moemm.md

Round 25 blitz: mm_down3 Q2_0 64-elem block step + run_q<2/6/10/14> (-11.8% in kprof, 361.9->319.2 ms), gdn_post gate/gamma preload (-0.7%), mm_gu3 half4 HS store pointer hoist

argon_r26_moemm.md

Round 26 blitz: qsa_attn_mm 32-bit ko/vo + float4 loads (-23.3..-29.9%), gemv IQ4_XS packed_uint4 5-reg codes (-1.5..-5% cold DRAM), DW<IQ4_XS/IQ4_NL/Q6_K> unroll + MM_V3_ONLY fast compile

argon_r27_gdn.md

Round 27 blitz: gdn_step_t 5->4 barriers + dt/ssm_a hoist + red2[IB-1] (-3.0..-7.9%), qsa_merge2/qsa_merge float2(m,l) + simd_max (-8..-20%), gemv_tile n_out=512 32,2,4,1 (-3.2..-25.5%) + Q2_0 640->2560 8,2,4,1

argon_r28_ple.md

Round 28 blitz: mtp.fc_hidden [t][c] stream batching (-52.1% n=2 / -29.3% n=3), ple.metal tg_sum barrier halving (-1.5..-5.8%), sample.metal tg_argmax unroll + router_route_8

argon_r29_moe.md

Round 29 blitz: mm_gu3 T_IQ2_S scale caching (-86.2 ms / -10% in kprof) & T_IQ2_XXS byte loads (-6.5%), moe_route e512/res_mask (-20.6..-26.4%) + deq<T_IQ2_S/XXS> (-1.1..-6.9%)

argon_r30_gdn.md

Round 30 blitz: W4<T_Q2_0>::dec FP16 0x6400 exponent trick (-10.5% mm_down3 / -37 ms in kprof -> 610.75 tok/s GPU prefill) + gdn_pre/post float4 simd_sum

argon_r31_dq.md

Round 31 blitz: mm_gu3 IQ2_S/IQ2_XXS/IQ2_XS/IQ3_S tab+hh & shift (+6.8..+8.2% / -57.9 ms in kprof), moe_down/gemv Q2_0R 0x6400 half-mantissa decode (-7.6..-17.4%), dq_f16 IQ4 tab init & perm=4 inverse contiguous write

argon_r32_moe.md

Round 32 blitz: qsa_flash sp[4] 3->1 barrier & ko[j] offset + unroll (-4.0..-5.5% / -6.1 ms in kprof), mm_gu3 T_IQ2_S half4 HS store (-9.0 ms in kprof -> 3175.3 ms / 624.5 tok/s GPU prefill!), deq uint2 grid loads + deq_iq_packed (-> 408.75 tok/s prefill & 42.16 tok/s decode!)

argon_r33_gdn.md

Round 33 blitz: gdn_step_t kl split + pre-barrier3 state commit (-1.1..-4.1%), gr_norm hoist (-6..-11%), GrdUnit<T_Q8_0,2> packed char4 (-10..-16%) -> 628.51 tok/s GPU prefill, mtp_ms 124.20/193.49/323.71ms

argon_r34_moemm.md

Round 34 blitz: gemm_hc_down TN=96 tail (-5.0..-7.2%), grd_acc ROLL=true at KS=4 (-1.9..-4.6% on _r2), W8<T_IQ3_S>::block + half4 HS store + IQ2_XS sboff -> 432.78 tok/s 2K prefill, 42.27 tok/s decode, 25.04 tok/s @10.2K

argon_r35_mtp.md

Round 35 blitz: Wr/Wr2 router dq reuse (-47 dispatches/chunk), gemm_f16_16_96_2 Walbe (-13.6%), M%128<=64 64_128_4 (1.14x @ M=306), hc_up_tile row-tail trim (-2.5% @ T=289), mtp_rms_xprep (-37.5%) & mtp_embed (-21.8%) & mtp_slop_pen (-96.5%) -> 629.05 tok/s GPU prefill, 25.26 tok/s @10.2K decode, mtp_ms 192.17/355.15ms

argon_r36_qsa.md

Round 36 blitz: predL_/predL2_ token-0 fast path when PREFETCH_DRAFT<=0 (-16.0% hot / -6.8% cold), TM>32 row-tail trim in gemm_hc_down (-4.7% @ T=1025) & gemm_swiglu_up (-1.4% @ T=1025, -0.8% @ T=1987) -> mtp_ms 190.62ms, 42.26 tok/s decode

argon_r5_scratch.md

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)

argon_r6_gr_gdn.md

Round 6 blitz: qsa_flash -20%, ple_moe_gate_norm_c + gated elim, -286 MB PLE/bo/yr aliasing (1.29 GB scratch), barrier-free gdn_pre/post, gr_norm_h* rv (366.8 tok/s prefill)

argon_r7_gdn.md

Round 7 blitz: single-pass simd_broadcast gdn_scanr_4 (-22.2%), register gemm_hc_up (-8.9%), ple_hist_adv + 1.13 GB scratch (382.0 tok/s prefill, 40.4 tok/s decode)

Round 8 blitz: prompt_chunk sigs/cuts shared-expert+MoE-plan overlap, W8 IQ2_XS/IQ2_XXS opt, gr_norm_h_stream_out, dq_f16+act_f16 half4 (385.1 tok/s 2K prefill, 122.9 tok/s short prefill)

argon_r9_quant.md

Round 9 blitz: qsa_flash 19.9->7.6 KiB TG memory (4 TGs/core, -17.5% / -51.4 ms), TAB>1 barrier guards, IQ3_S dead sign load removal (392.02 tok/s 2K prefill, 40.11 tok/s decode)

astra_r38_qsa.md

Round 38: hold_region layer-0 allow_hot=false + prompt/decode bytes_read_/reads_ split + mtp_prefill tail_cap sync with MTP_WIN -> 35.47 tok/s @10.2K decode, 473.80 tok/s @10.2K prefill, 42.57 tok/s @2K decode, 37.20 tok/s reasoning

astra_r39_qsa.md

round 39: port serving fixes and PLE reuse; verify parity, record perf uncertainty

opus_audit_phase0_2_m5_v2.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase0_2_m5_v3.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase0_2_m5_v4.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase0_2_m5_v5.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase0_2_m5.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase1_methodology.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase1a_v2.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase1b.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase1c_v2.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase1c.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase3_local_v1.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_phase3_local_v2.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_audit_slice_replay.md

keep 00-..04- docs and non-persona engine/pruning files at top level; keep only persona/anti-slop in persona/

opus_colibri_scout.md

Add audits/opus_colibri_scout.md: Colibri vs Strata Metal audit + M5 hazard/wait/entropy benchmarks

opus_decode_fusion.md

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)

opus_gdn_qsa.md

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)

opus_mtp_turbo.md

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)

opus_ple_moe_out.md

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)

opus_prefill_opt.md

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

opus_r10_qsa.md

Round 10 blitz: qsa_flash 2.2x (-51.9% / -146.1 ms in kprof via unconditional hoisted rv/rk + vec4 epilogue), gdn_pre float4 conv, uint4 mm_gather_h + s_id-free mm_place

opus_r11_dense.md

Round 11 blitz: gemm_f16 tile selector (-18% / -234 ms), qsa_attn_mm hoisted rv/rk (-44.5%), branchless W8<T_IQ3_S> (-50.8% / 2.03x) + W8<T_IQ2_XS> k_hoist (-57 ms), 3,397.5 ms GPU kprof (583.7 tok/s GPU) & 2048 prompt chunk on :8080

opus_r12_moemm.md

Round 12 blitz: qsa_topk 3.85x (cached keys + simdgroup scan + paired emit), gemm_hc_up_reg_32_128_4 + unrolled epilogue (-6.9%), Q6_K hoisted mm_gu3 (-4.3%), 3,375.4 ms GPU kprof (587.5 tok/s GPU)

opus_r13_hcdown.md

Round 13 blitz: gemm_hc_down + gemm_swiglu_up K4 register epilogues (-144 dispatches, -985 MB DRAM/chunk, 600.6 tok/s GPU kprof), dq_f16 exact-simdgroup 2D grid, SpecGate + mtp_draft PLE prefetch

opus_r14_moemm.md

Round 14 blitz: qsa_flash/qsa_attn_mm byte-offset + zero-scale hoist (-9.7% at pos0>=2048, 443.2 tok/s on 10.18K Albedo prompt), gdn_scanr_4 hk-locality, compacted moe_comb roff, mm_gu3 IQ2_XS half4 HS store

opus_r15_win.md

Round 15 blitz: prompt_chunk router+plan/shexp wave zipping (-190 barriers), fused Wal+Wbe N=96 GEMM (-36 dispatches), QSA PSO pre-warm (-202 ms Chunk-1 stall), qsa_score_mm causal-future skip (-19.4%), qsa_prep 320B TG + 2 barriers

opus_r16_kvik.md

Round 16 blitz: fused QSA Wk+Wv+Wik(+Wiq) single GEMM (-24..-36 dispatches, -5.6 MB scratch), decode resident-MoE + shared-expert wave zip (-48 barriers), gemv TAB>1 guards, ple_moe_gate_norm_c SIMD compaction + ple_conv half4/t>=9

opus_r17_mtp.md

Round 17 blitz: mtp.metal float4 vectorization + Q8_0 gemv_catk, barrier-free simd_shuffle_down in gr_up_f/gr_up3/qsa_merge, gdn_step ib==0 readout pruning + qsa_attn char4 K loads (39.56 tok/s warm decode)

opus_r18_moe.md

Round 18 blitz: per-stream gr_norm_comb 5.09x (-80%), xprep float4 (-46.8%), moe_combine_sh (-49%), Q6_K moe_gu NE>=3 (-25%), qsa_topk2/qsa_score (-5..9%) -> 40.03 tok/s decode & 445.79 tok/s 10.18K prefill

opus_r19_prewarm.md

Round 19 blitz: gemv NTOK/NR/KS unroll (-5.4% to -29.4% at ntok>=2), mtp_pick/slop unroll_count(4) (-27..60%), ple_comb_query_gate_norm + load-time prewarm -> 40.44 tok/s decode & 212.5 ms mtp_ms

opus_r20_mtp_qsa.md

Round 20 blitz: mtp_load mlock 708 MB experts (mtp_ms 398->122 ms on 11K, 433->202 ms on 2K; 40.57 tok/s decode), qsa_attn (-20..40%) + qsa_merge (-10..13%), gdn_scanr_4 32-bit offsets (186.15 ms kprof)

Round 21 blitz: gr_down/gr_up GR_NTOK unroll (-5.4..18.0%), qsa_attn2/qsa_merge2 (-36.8..40.6% / -8.3..14.6%), gdn_step st_off + moe_plan topk10 unroll_count(10) (-9.5..26.6%) -> 41.66 tok/s decode & 193.9 ms mtp_ms

Round 22 blitz: gr_down_4_4/_r2 NR=2 xn reuse (-8.1..26.1%), moe_down/moe_gu q2_combine (-7.7..22.5%) + plan_body simd_prefix (-14.6..35.4%), swiglu_prep barrier-free -> 35.66 tok/s coding & 24.71 tok/s @10.2K decode

opus_r23_quant.md

Round 23 blitz: ksg popcount + gemv_direct_x/gemv_q2_raw_combine + mm_gu3 sign4_iq2xs/ptr-folding -> 42.01 tok/s decode & 602.1 tok/s GPU prefill

opus_r24_prompt.md

Round 24 blitz: W8<T_Q6_K> (+12.8..24.9%), qsa_flash 32-bit ko/vo (-11.7% @8K), TAB 1024/512 + gr_norm_h_moec k-outer + gemm_hc_down *0.25f

opus_r25_gdn.md

Round 25 blitz: mm_down3 Q2_0 64-elem block step + run_q<2/6/10/14> (-11.8% in kprof, 361.9->319.2 ms), gdn_post gate/gamma preload (-0.7%), mm_gu3 half4 HS store pointer hoist

opus_r26_gemv.md

Round 26 blitz: qsa_attn_mm 32-bit ko/vo + float4 loads (-23.3..-29.9%), gemv IQ4_XS packed_uint4 5-reg codes (-1.5..-5% cold DRAM), DW<IQ4_XS/IQ4_NL/Q6_K> unroll + MM_V3_ONLY fast compile

opus_r27_gemv.md

Round 27 blitz: gdn_step_t 5->4 barriers + dt/ssm_a hoist + red2[IB-1] (-3.0..-7.9%), qsa_merge2/qsa_merge float2(m,l) + simd_max (-8..-20%), gemv_tile n_out=512 32,2,4,1 (-3.2..-25.5%) + Q2_0 640->2560 8,2,4,1

opus_r28_mtp.md

Round 28 blitz: mtp.fc_hidden [t][c] stream batching (-52.1% n=2 / -29.3% n=3), ple.metal tg_sum barrier halving (-1.5..-5.8%), sample.metal tg_argmax unroll + router_route_8

opus_r29_prompt.md

Round 29 blitz: mm_gu3 T_IQ2_S scale caching (-86.2 ms / -10% in kprof) & T_IQ2_XXS byte loads (-6.5%), moe_route e512/res_mask (-20.6..-26.4%) + deq<T_IQ2_S/XXS> (-1.1..-6.9%)

opus_r30_moemm.md

Round 30 blitz: W4<T_Q2_0>::dec FP16 0x6400 exponent trick (-10.5% mm_down3 / -37 ms in kprof -> 610.75 tok/s GPU prefill) + gdn_pre/post float4 simd_sum

opus_r31_moemm.md

Round 31 blitz: mm_gu3 IQ2_S/IQ2_XXS/IQ2_XS/IQ3_S tab+hh & shift (+6.8..+8.2% / -57.9 ms in kprof), moe_down/gemv Q2_0R 0x6400 half-mantissa decode (-7.6..-17.4%), dq_f16 IQ4 tab init & perm=4 inverse contiguous write

opus_r32_moemm.md

Round 32 blitz: qsa_flash sp[4] 3->1 barrier & ko[j] offset + unroll (-4.0..-5.5% / -6.1 ms in kprof), mm_gu3 T_IQ2_S half4 HS store (-9.0 ms in kprof -> 3175.3 ms / 624.5 tok/s GPU prefill!), deq uint2 grid loads + deq_iq_packed (-> 408.75 tok/s prefill & 42.16 tok/s decode!)

Round 33 blitz: gdn_step_t kl split + pre-barrier3 state commit (-1.1..-4.1%), gr_norm hoist (-6..-11%), GrdUnit<T_Q8_0,2> packed char4 (-10..-16%) -> 628.51 tok/s GPU prefill, mtp_ms 124.20/193.49/323.71ms

Round 34 blitz: gemm_hc_down TN=96 tail (-5.0..-7.2%), grd_acc ROLL=true at KS=4 (-1.9..-4.6% on _r2), W8<T_IQ3_S>::block + half4 HS store + IQ2_XS sboff -> 432.78 tok/s 2K prefill, 42.27 tok/s decode, 25.04 tok/s @10.2K

opus_r35_dense.md

Round 35 blitz: Wr/Wr2 router dq reuse (-47 dispatches/chunk), gemm_f16_16_96_2 Walbe (-13.6%), M%128<=64 64_128_4 (1.14x @ M=306), hc_up_tile row-tail trim (-2.5% @ T=289), mtp_rms_xprep (-37.5%) & mtp_embed (-21.8%) & mtp_slop_pen (-96.5%) -> 629.05 tok/s GPU prefill, 25.26 tok/s @10.2K decode, mtp_ms 192.17/355.15ms

opus_r36_decode.md

Round 36 blitz: predL_/predL2_ token-0 fast path when PREFETCH_DRAFT<=0 (-16.0% hot / -6.8% cold), TM>32 row-tail trim in gemm_hc_down (-4.7% @ T=1025) & gemm_swiglu_up (-1.4% @ T=1025, -0.8% @ T=1987) -> mtp_ms 190.62ms, 42.26 tok/s decode

opus_r5_comb.md

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 6 blitz: qsa_flash -20%, ple_moe_gate_norm_c + gated elim, -286 MB PLE/bo/yr aliasing (1.29 GB scratch), barrier-free gdn_pre/post, gr_norm_h* rv (366.8 tok/s prefill)

opus_r7_scratch.md

Round 7 blitz: single-pass simd_broadcast gdn_scanr_4 (-22.2%), register gemm_hc_up (-8.9%), ple_hist_adv + 1.13 GB scratch (382.0 tok/s prefill, 40.4 tok/s decode)

opus_r8_moemm.md

Round 8 blitz: prompt_chunk sigs/cuts shared-expert+MoE-plan overlap, W8 IQ2_XS/IQ2_XXS opt, gr_norm_h_stream_out, dq_f16+act_f16 half4 (385.1 tok/s 2K prefill, 122.9 tok/s short prefill)

Round 9 blitz: qsa_flash 19.9->7.6 KiB TG memory (4 TGs/core, -17.5% / -51.4 ms), TAB>1 barrier guards, IQ3_S dead sign load removal (392.02 tok/s 2K prefill, 40.11 tok/s decode)

opus_strata_metal_opts_v1.md

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)

opus_strata_metal_opts_v2.md

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)

opus_strata_metal_opts_v3.md

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)

r39_integration.md

document paired-run review limits and baseline provenance

r39_opus_review.md

document paired-run review limits and baseline provenance

round 39: port serving fixes and PLE reuse; verify parity, record perf uncertainty

r40_opus_readme_review.md

round 40: add apple silicon README, setup script, and complete strata metal diff

r40_rocm14_metal_optimizations.md

round 43: port rocm adaptive-kv-stream opts to Metal (WY GDN prefill, MTP defer_catchup, GDN ring rollback, 1-sg qsa_prep, delayed-softmax topk, coalesced lazy mmap)

sol_async_stream_prefill.md

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)

sol_moce_spec.md

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)

sol_moe_cache_policy.md

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

sol_mtp_fuse.md

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 10 blitz: qsa_flash 2.2x (-51.9% / -146.1 ms in kprof via unconditional hoisted rv/rk + vec4 epilogue), gdn_pre float4 conv, uint4 mm_gather_h + s_id-free mm_place

Round 11 blitz: gemm_f16 tile selector (-18% / -234 ms), qsa_attn_mm hoisted rv/rk (-44.5%), branchless W8<T_IQ3_S> (-50.8% / 2.03x) + W8<T_IQ2_XS> k_hoist (-57 ms), 3,397.5 ms GPU kprof (583.7 tok/s GPU) & 2048 prompt chunk on :8080

Round 12 blitz: qsa_topk 3.85x (cached keys + simdgroup scan + paired emit), gemm_hc_up_reg_32_128_4 + unrolled epilogue (-6.9%), Q6_K hoisted mm_gu3 (-4.3%), 3,375.4 ms GPU kprof (587.5 tok/s GPU)

sol_r13_specgate.md

Round 13 blitz: gemm_hc_down + gemm_swiglu_up K4 register epilogues (-144 dispatches, -985 MB DRAM/chunk, 600.6 tok/s GPU kprof), dq_f16 exact-simdgroup 2D grid, SpecGate + mtp_draft PLE prefetch

Round 14 blitz: qsa_flash/qsa_attn_mm byte-offset + zero-scale hoist (-9.7% at pos0>=2048, 443.2 tok/s on 10.18K Albedo prompt), gdn_scanr_4 hk-locality, compacted moe_comb roff, mm_gu3 IQ2_XS half4 HS store

Round 15 blitz: prompt_chunk router+plan/shexp wave zipping (-190 barriers), fused Wal+Wbe N=96 GEMM (-36 dispatches), QSA PSO pre-warm (-202 ms Chunk-1 stall), qsa_score_mm causal-future skip (-19.4%), qsa_prep 320B TG + 2 barriers

Round 16 blitz: fused QSA Wk+Wv+Wik(+Wiq) single GEMM (-24..-36 dispatches, -5.6 MB scratch), decode resident-MoE + shared-expert wave zip (-48 barriers), gemv TAB>1 guards, ple_moe_gate_norm_c SIMD compaction + ple_conv half4/t>=9

sol_r17_shuf.md

Round 17 blitz: mtp.metal float4 vectorization + Q8_0 gemv_catk, barrier-free simd_shuffle_down in gr_up_f/gr_up3/qsa_merge, gdn_step ib==0 readout pruning + qsa_attn char4 K loads (39.56 tok/s warm decode)

sol_r18_grcomb.md

Round 18 blitz: per-stream gr_norm_comb 5.09x (-80%), xprep float4 (-46.8%), moe_combine_sh (-49%), Q6_K moe_gu NE>=3 (-25%), qsa_topk2/qsa_score (-5..9%) -> 40.03 tok/s decode & 445.79 tok/s 10.18K prefill

sol_r19_sample.md

Round 19 blitz: gemv NTOK/NR/KS unroll (-5.4% to -29.4% at ntok>=2), mtp_pick/slop unroll_count(4) (-27..60%), ple_comb_query_gate_norm + load-time prewarm -> 40.44 tok/s decode & 212.5 ms mtp_ms

sol_r20_gather.md

Round 20 blitz: mtp_load mlock 708 MB experts (mtp_ms 398->122 ms on 11K, 433->202 ms on 2K; 40.57 tok/s decode), qsa_attn (-20..40%) + qsa_merge (-10..13%), gdn_scanr_4 32-bit offsets (186.15 ms kprof)

sol_r21_qsa2.md

Round 21 blitz: gr_down/gr_up GR_NTOK unroll (-5.4..18.0%), qsa_attn2/qsa_merge2 (-36.8..40.6% / -8.3..14.6%), gdn_step st_off + moe_plan topk10 unroll_count(10) (-9.5..26.6%) -> 41.66 tok/s decode & 193.9 ms mtp_ms

sol_r22_router_glue.md

Round 22 blitz: gr_down_4_4/_r2 NR=2 xn reuse (-8.1..26.1%), moe_down/moe_gu q2_combine (-7.7..22.5%) + plan_body simd_prefix (-14.6..35.4%), swiglu_prep barrier-free -> 35.66 tok/s coding & 24.71 tok/s @10.2K decode

sol_r23_gemv.md

Round 23 blitz: ksg popcount + gemv_direct_x/gemv_q2_raw_combine + mm_gu3 sign4_iq2xs/ptr-folding -> 42.01 tok/s decode & 602.1 tok/s GPU prefill

Round 24 blitz: W8<T_Q6_K> (+12.8..24.9%), qsa_flash 32-bit ko/vo (-11.7% @8K), TAB 1024/512 + gr_norm_h_moec k-outer + gemm_hc_down *0.25f

Round 25 blitz: mm_down3 Q2_0 64-elem block step + run_q<2/6/10/14> (-11.8% in kprof, 361.9->319.2 ms), gdn_post gate/gamma preload (-0.7%), mm_gu3 half4 HS store pointer hoist

Round 26 blitz: qsa_attn_mm 32-bit ko/vo + float4 loads (-23.3..-29.9%), gemv IQ4_XS packed_uint4 5-reg codes (-1.5..-5% cold DRAM), DW<IQ4_XS/IQ4_NL/Q6_K> unroll + MM_V3_ONLY fast compile

Round 27 blitz: gdn_step_t 5->4 barriers + dt/ssm_a hoist + red2[IB-1] (-3.0..-7.9%), qsa_merge2/qsa_merge float2(m,l) + simd_max (-8..-20%), gemv_tile n_out=512 32,2,4,1 (-3.2..-25.5%) + Q2_0 640->2560 8,2,4,1

sol_r28_router.md

Round 28 blitz: mtp.fc_hidden [t][c] stream batching (-52.1% n=2 / -29.3% n=3), ple.metal tg_sum barrier halving (-1.5..-5.8%), sample.metal tg_argmax unroll + router_route_8

sol_r29_moemm.md

Round 29 blitz: mm_gu3 T_IQ2_S scale caching (-86.2 ms / -10% in kprof) & T_IQ2_XXS byte loads (-6.5%), moe_route e512/res_mask (-20.6..-26.4%) + deq<T_IQ2_S/XXS> (-1.1..-6.9%)

sol_r30_prompt.md

Round 30 blitz: W4<T_Q2_0>::dec FP16 0x6400 exponent trick (-10.5% mm_down3 / -37 ms in kprof -> 610.75 tok/s GPU prefill) + gdn_pre/post float4 simd_sum

Round 31 blitz: mm_gu3 IQ2_S/IQ2_XXS/IQ2_XS/IQ3_S tab+hh & shift (+6.8..+8.2% / -57.9 ms in kprof), moe_down/gemv Q2_0R 0x6400 half-mantissa decode (-7.6..-17.4%), dq_f16 IQ4 tab init & perm=4 inverse contiguous write

Round 32 blitz: qsa_flash sp[4] 3->1 barrier & ko[j] offset + unroll (-4.0..-5.5% / -6.1 ms in kprof), mm_gu3 T_IQ2_S half4 HS store (-9.0 ms in kprof -> 3175.3 ms / 624.5 tok/s GPU prefill!), deq uint2 grid loads + deq_iq_packed (-> 408.75 tok/s prefill & 42.16 tok/s decode!)

Round 33 blitz: gdn_step_t kl split + pre-barrier3 state commit (-1.1..-4.1%), gr_norm hoist (-6..-11%), GrdUnit<T_Q8_0,2> packed char4 (-10..-16%) -> 628.51 tok/s GPU prefill, mtp_ms 124.20/193.49/323.71ms

sol_r34_prompt.md

Round 34 blitz: gemm_hc_down TN=96 tail (-5.0..-7.2%), grd_acc ROLL=true at KS=4 (-1.9..-4.6% on _r2), W8<T_IQ3_S>::block + half4 HS store + IQ2_XS sboff -> 432.78 tok/s 2K prefill, 42.27 tok/s decode, 25.04 tok/s @10.2K

sol_r35_prompt.md

Round 35 blitz: Wr/Wr2 router dq reuse (-47 dispatches/chunk), gemm_f16_16_96_2 Walbe (-13.6%), M%128<=64 64_128_4 (1.14x @ M=306), hc_up_tile row-tail trim (-2.5% @ T=289), mtp_rms_xprep (-37.5%) & mtp_embed (-21.8%) & mtp_slop_pen (-96.5%) -> 629.05 tok/s GPU prefill, 25.26 tok/s @10.2K decode, mtp_ms 192.17/355.15ms

sol_r36_prompt.md

Round 36 blitz: predL_/predL2_ token-0 fast path when PREFETCH_DRAFT<=0 (-16.0% hot / -6.8% cold), TM>32 row-tail trim in gemm_hc_down (-4.7% @ T=1025) & gemm_swiglu_up (-1.4% @ T=1025, -0.8% @ T=1987) -> mtp_ms 190.62ms, 42.26 tok/s decode

sol_r5_mtp_qsa.md

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 6 blitz: qsa_flash -20%, ple_moe_gate_norm_c + gated elim, -286 MB PLE/bo/yr aliasing (1.29 GB scratch), barrier-free gdn_pre/post, gr_norm_h* rv (366.8 tok/s prefill)

Round 7 blitz: single-pass simd_broadcast gdn_scanr_4 (-22.2%), register gemm_hc_up (-8.9%), ple_hist_adv + 1.13 GB scratch (382.0 tok/s prefill, 40.4 tok/s decode)

sol_r8_grvec.md

Round 8 blitz: prompt_chunk sigs/cuts shared-expert+MoE-plan overlap, W8 IQ2_XS/IQ2_XXS opt, gr_norm_h_stream_out, dq_f16+act_f16 half4 (385.1 tok/s 2K prefill, 122.9 tok/s short prefill)

sol_r9_moemm.md

Round 9 blitz: qsa_flash 19.9->7.6 KiB TG memory (4 TGs/core, -17.5% / -51.4 ms), TAB>1 barrier guards, IQ3_S dead sign load removal (392.02 tok/s 2K prefill, 40.11 tok/s decode)

sol_slop4_mtp.md

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)

sol_spec_mtp_decode.md

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