fpga-archive-reports-record · EAGLE4 FPGA archive
Raising K_C above 64 on the 7-card set (COMBO3 0xE4B713D6 x4 + M1361 0xE4B71361 x2): hardware support, price, and the K_C window plan (2026-11-16)
Program 1: the 7-card Llama-3.2-3B 2-bit EAGLE4 drafter on talon2 (ATT TP4 COMBO3 on 17/3d/bd/e1, MLPV M1361 e24deb3f on 50/63, H200 verify).
Motto: a track does not end by denying the idea. "Enable higher candidate count to get more acc" (user, 2026-11-16) is treated as an engineering problem. Below are the measured price and the window that measures the rest.
Scope (coordinator, 2026-11-16): the depth x width grid at K_C 64 is ALREADY RUNNING on talon2 (box1-two-v2: runs/acwin_20261116b/dw_sweep/, doc DW_SWEEP_7CARD_COMBO3_20261116.md). This doc covers the K_C > 64 part. Its D x W content is limited to what was measured before the grid and how to pick the K_C cells from the grid. Section 4 part A is kept only as the reference ROUNDS for cells the grid may not cover.
Research worker for box1-two-v2. Read-only plus create-only. talon2 was not touched (no ssh to talon2, no card, no lock). buildx04 got one read-only grep. Nothing was flashed, killed, moved or removed.
Tags: [M] measured (the artifact is named) · [D] derived / modelled. Every [D] number is in the first section.
Calculations and hypothesis (every projection in this doc is here)
-
D1. Base. COMBO3 W2 d11 = 961.43 tok/s, 2,839 steps, 19,615 tokens. That is 20.40 s per run, or 7.186 ms/step.
-
D2. K_C cost per verify step, from the 0x13BC + M1361 window of 11-12 (
accept_gap_20261112/silicon/serve_ag/SUMMARY.tsvmd5 26bb71b9: time_s / steps).- The reference is 7.436 ms/step.
- K_C 128 adds +0.179 ms.
- K_C 256 adds +0.863 ms.
- With TREE_STOP off, K_C 256 vs 64 adds +1.222 ms.
- Per W=2 item (7.40 / 7.32 forwarded levels x 2 items): +12 us at K_C 128 and +59 us at K_C 256 (about 800 / 3,890 cycles at 66 MHz).
- The 64 -> 128 step is cheap because the G8 window (2,969 cycles per item at K_C 64) is still mostly hidden under TREE (2,224 cycles per item; M1361 census,
TREE4_SILICON_RCA_20261114.md:73-76). Past about 128 every extra candidate is exposed. - Marginal slope 128 -> 256: 5.34 us per candidate per step = about 24 cycles per candidate per item.
-
D3. K_C 512 has never run on silicon. Extrapolating that slope gives +2.23 ms/step. The G8 cycle law gives +2.0 ms. Range used: +2.0 to +2.2 ms/step.
-
D4. Steps.
- At TREE_STOP on and theta -2.5 [M]: K_C 128 / 256 = 2,851 / 2,857.
- K_C 512 at theta -2.5: about 2,860 [D].
- With theta re-tuned so the stop pattern matches K_C 64 (D5): about 2,836. This uses the stop-off K_C gain of -3 steps [M: tz1 2,806 -> kz1 2,803].
-
D5. Theta re-tune per K_C. A larger K_C lowers every per-parent softmax.
- Measured shift on silicon: -0.020 bits per level at K_C 256 (
DESIGN_FABLE_ACCEPT_NOLEVEL_20261112.md:275-283). - The replay maps theta -2.65 at K_C 256 back to the K_C 64 / theta -2.5 stop pattern (same doc, S3).
- Derived values: K_C 128 -> theta -2.58 (stop share 0.5641 vs 0.5525 => about 0.01 bit per level). K_C 512 -> theta -2.72 (extrapolated 0.025-0.03 bit per level).
- Measured shift on silicon: -0.020 bits per level at K_C 256 (
-
D6. Modelled tok/s on COMBO3, W2 d11, = 19,615 / (steps x (7.186 + delta) ms):
K_C delta ms/step TS1, theta -2.5 TS1, theta re-tuned accept (TS1 -2.5 / re-tuned / twin ceiling) 64 0 [M] 961.4 [M] — 6.91 [M] 128 +0.179 [M on 0x13BC set] 934.1 (-2.8 %) 939.0 (-2.3 %) 6.88 [M] / ~6.92 / — 256 +0.863 [M on 0x13BC set] 852.9 (-11.3 %) 859.2 (-10.6 %) 6.87 [M] / ~6.92 / 6.95 512 +2.0..2.2 [D3] 728-747 (-22..-24 %) 734-753 (-22..-24 %) ~6.86 / ~6.92 / ? - "Twin ceiling" is the extra the silicon would gain if it matched its GPU twin: 15 steps, 2,839 -> 2,824 = 6.95.
- COMBO3 makes the ATT side faster, which leaves the MLPV G8 more exposed. The deltas could therefore be LARGER on COMBO3. The sweep measures them.
-
D7. Break-even. One verify step = 0.035 % tok/s.
- K_C 128 must save about 80 steps (accept >= 7.11) to pay for its +2.4 % ms/step. That is above the whole 63-step gap to GPU-only (7.05).
- On these images no K_C > 64 can win on tok/s, even if it closed the entire accept gap. It is an accuracy-for-speed trade, and that is a USER ruling (as K_C 32 was).
-
D8. Hypothesis for the open finding. On the GPU twin, K_C 256 recovers 15 of the card head's 18 steps. On silicon it recovers 3. Candidate causes:
- the card's cheap-stream prefilters (the vs_link parity merge prefilter, the tiler threshold);
- or the bridge bound:
kc_en= ACCFIX bit2. The K_C-uniform denominator shift seen on silicon implies kc_en is on, but 0x15C has not been read on the MLPVs. - Neither has been measured. B0 below discriminates them with GPU time only.
-
D9. Wall time. Rounds measured on this kit: 65 s per round at W2 d11, 69 s at W4 (start-to-start, box1
/var/tmp/box1two_scratch_20261116/d6_w2.log,d6_w4.log). Assumed 70-75 s for K_C 512.- K_C part only (B0-B5): 28 serve rounds + 6 K_C goldens + 4 GPU-only twin cells = about 45 min warm (right after the grid, same boot, T4_GOLDEN_OK present).
- +25 min per hang recovery (cold AC + bring-up). Budget 1 h 15 min.
- Part A, had it been needed: 42 rounds, about 50 min.
0. Answer first
-
Every K_C from 1 to 512 is a runtime register on the current images. No rebuild is needed.
- MLPV
0x150 kc_cfgdefaults to 0, which means the compile K_C,EAGLE4_G8_K_C 512. - Values above 512 silently clamp to 512.
- The ATT image COMBO3 never reads K_C.
- K_C 128 and 256 already ran clean on M1361 [M]. K_C 512 has xsim arms MATCH on the G8_OWN source but has never run on silicon.
- MLPV
-
Prior K_C > 64 data on 7 cards exists and is negative on tok/s [M, 0x13BC x4 + M1361, d11 W2, 11-12]. With TREE_STOP on, accept FALLS:
K_C tok/s accept steps texts 128 903.45 (-2.78 %) 6.88 2,851 record 1b0779cf 256 827.13 (-11.0 %) 6.87 2,857 1b1e44c1 256, TS off 760.59 (-18.1 %) 7.00 2,803 (-3 vs K_C 64 TS off) - Cause, measured: the per-parent softmax runs over K_C candidates. It shifts scores by -0.020 bits per level, so TREE_STOP fires earlier.
- K_C 32 bought +1.19 % for -0.015 accept.
-
Shortlist share of the 6.91 vs 7.05 gap (63 steps) [M]:
term steps share TREE_STOP 33 52 % card cheap head x K_C 64 (GPU twin) 18 29 % 3 text-divergent requests 12 19 % - With the fp head, K_C makes no difference: GPU K_C 64 = 128 = 256 = 2,776 steps.
- On the GPU twin, the card head at K_C 256 recovers 15 of its 18 steps. On silicon, K_C 256 recovers only 3. That gap is the open item. B0 below is its cheapest probe: GPU-only twin captures plus
kc_cap_compare.pyon the captures already on disk.
-
Cost. See D6: about -2.8 % / -11 % / -23 % tok/s at K_C 128 / 256 / 512.
-
The K_C sweep runs on the current images with host env only: K_C 128 / 256 / 512 are runtime-settable (MLPV 0x150 via
EAGLE4_TREE_KC= ROUNDS field 6), with no flash and no rebuild. It can be queued right after the D x W grid on the same boot. I added three create-only kit files:kcdw_serve_w.sh/kcdw_steps.sh, which add per-round K_C, TREE_STOP, width, VTOK and theta fields to the frozen ATT-combo window scripts, a post-round K_C check and akcgolden <kc>phase;kcdw_mk_head512.shfor the twin arm.
No build spec is needed below K_C 512. Section 6 lists what would make a higher K_C pay.
1. What the images support [M of the RTL / host, file:line]
Fork provenance (derived from the stage records, not memory):
- MLPV M1361:
/mnt/models/stage_E4B71361/STAGE_SHA_E4B71361.txt, forkforks/mlpv_g8_own_20261111v3 on parentfourcard_4c1_mlpv_tp4_20261109(the 0x134D source).- Staged tree:
/mnt/data/eagle4_scratch/mlpv_g8own_20261111/4c1tp4_1361_g8own_mlpv_stage/forks/tail_fgi_20260930/src_full. - md5s: drafter 845ca21c, config c7c25ed2, rescore 2cb84fc7, own_link ffd8d5b9, bridge 2768551a, control_regs 2172ed4d. Deriver
scripts/4c1/hedge/derive_1361_g8own.py.
- Staged tree:
- ATT COMBO3:
combo3_20261113/DERIVATION_COMBO3.txt(8e15bb02). Stagelutg_d9_tp4stk_D6_stagefrom the 0x13BC BC stage, forkforks/head_hxf_cons/src_full, = 0x13BC + G5_P2W 8 + R3_EARLY 1 + KV1H2C 1 + HANDOFF_CUT 1.
| item | value | evidence |
|---|---|---|
| compile K_C (max) | 512 | eagle4_config.svh:875-876 EAGLE4_G8_K_C 512 |
| runtime K_C register | MLPV 0x150 kc_cfg[15:0], reset 0 (= 512), read-back |
control_regs.sv:177,693,923,1164; eagle4_drafter.sv:1239-1241 |
| clamp | kc_eff = (kc_cfg == 0 || kc_cfg > K_C) ? K_C : kc_cfg, sampled per item at G8 launch (kcq <= kc_eff) |
g8_rescore_p.sv:542-543,1604 |
| G8 buffers | sel_score / sel_idx [0:K_C-1], rescored[] sized at K_C 512; the FG8 frontier has n_rows [$clog2(K_C):0] |
g8_rescore_p.sv:548-549; g8_spec_frontend.sv:60,82 |
| softmax bridge bound | follows kcq only when accfix_kc_en (= ACCFIX_KC_P 1 AND 0x15C bit2); otherwise it sweeps all 512 |
tree_softmax_bridge.sv:374-375,621; eagle4_drafter.sv:1249,5483; eagle4_config.svh:937-938 |
| G8_OWN link payload | ceil(kcq/7) score beats per item (NE 7, WIN 32): 10 / 19 / 37 / 74 beats at K_C 64 / 128 / 256 / 512; kcq field [$clog2(K_C):0] = 10 bits |
mlpv_g8_own_link.sv:14,33,49,60 |
| both MLPVs must hold the same K_C | unequal kcq misaligns the score beats with the next G9 exchange (audit M1) | mlpv_g8_own_link.sv:14-15; host writes both, c4_loop.py:354-355, c4_hx7.py:471-472 |
| ATT side | G8_* straps are inert on PIPE_ROLE 1 (module absent from the ATT netlist); the host writes 0x150 only to MLPVs | ATT_CARD_CONSOLIDATION_20261109.md:357; c4_loop.py:354 |
| host knob | EAGLE4_TREE_KC=$KC (runner KC=${KC:-64}); window kit ROUNDS field 6 |
scripts/4c1/run_4c1_sglang.sh:36,99; attcombo_20261116/ship/ac_serve_w.sh:19,37,49 |
| xsim at K_C > 64, G8_OWN source | 128 / 256 (plain, stressed LAT 200 JIT 40 SKEW 300, LAT 0): 8/8 walks MATCH, slot / rescore / frontier mismatches 0, beats 19 / 37 both cards, own_sum = K_C; K_C 512 MATCH (11-11) | buildx04 /shared/scratch/mlpv_g8_own_20261111/arms/k4_*/vs.out (read 2026-11-16, XSIM_EXIT 0, end 00:15-00:18Z 11-13); AUDIT_MLPV_G8_OWN_20261111.md:133,262 |
| tree width | RTL TREE_WIDTH_MAX 8, runtime 0x1CC tree_width[3:0] 1..8, latched per tree |
eagle4_config.svh:821-822; control_regs.sv:835,1052 |
| host width gate | W > 2 refused (rc 61) unless EAGLE4_4C1_ALLOW_WIDE=1 |
run_4c1_sglang.sh:48-49 |
| card-side TREE_STOP verdict | only at W == 2: at any other width, W=1 included, every level waits for a host verdict over PCIe | common/rtl/loop_seq_d2.sv:516,528 (stage copy b086b6b0) |
| verify tokens | VTOK must be D+1 at W=1: SGLang refused V20 with a 12-node tree (rc 247, can not run cuda graph ... verify_num = 12); W>=2 at d>=7 has >= 27 nodes, so V20 is fine |
DESIGN_FABLE_ACCEPT_NOLEVEL_20261112.md S2 row wc1 |
Unknown to close (one read-only MMIO): MLPV 0x15C bit2 (kc_en). [D] It is on, because the K_C-uniform score shift seen on silicon needs the bridge to normalise over kcq. Read 0x15C on 50/63 with c4 status or read before B.
2. Measured cells already in hand (do not re-run) [M]
| set | cell (humaneval 80x256, fa3 fg4 cp1, V20) | tok/s | accept | steps | texts | source |
|---|---|---|---|---|---|---|
| COMBO3 boot 79cc9d1b | W2 d11 TS1 K_C 64 | 960.23 / 960.97 / 963.10 (KV1H2C off 955.22) | 6.91 | 2,839 | 1b0779cf | ATT_COMBO_SILICON_20261116.md; box1 /var/tmp/box1two_scratch_20261116/d6_w2.log, d6_off.log |
| COMBO3 | W2 d15 TS1 (smoke, 1 round) | 959.34 | 7.17 | 2,735 | 3c26f0b0 | d6_smoke.log |
| COMBO3 | W4 d11 TS0 | 792.67 / 792.32 | 7.20 | 2,715 | 54f88d49 | d6_w4.log |
| D5F 0x13DB (same MLPV) | W2 d11 / d15 / W4 d11 TS0 | 942.53 / 942.49 / 782.22 | 6.91 / 7.17 / 7.20 | same texts as COMBO3 | db_*.log |
|
| 0x13BE + M1361 | W2 d11 TS1 | 938.94 | 6.91 | 2,839 | 1b0779cf | tracker 11-13 23:5x |
| 0x13BE + M1361 | W2 d11 TS0 | 887.36 | 6.99 | 2,806 | c1422a7e | W4_SERVE_BOUND_20261115.md:213 |
| 0x13BE + M1361 | W3 d11 TS1 | 789.58 / 790.92 | 7.18 | tracker 11-14 12:07 (4C1_DELIVERY_TRACKER.md:1735) |
||
| 0x13BE + M1361 | W4 d11 TS1 / TS0 | 714.59 / 756.05 | 7.19 / 7.20 | 2,721 / 2,715 | 54f88d49 | W4_SERVE_BOUND_20261115.md:210-211 |
| 0x13BE + M1361 | W8 d11 TS1 | 475.83 / 476.66 | 7.21 | tracker :1735 | ||
| 0x13BE + M1361 | W2 d15 TS1 | 935.83 | 7.17 | tracker 11-14 02:45 | ||
| 0x13BC + M1361 | W2 d9 TS1 | 894.47 / 891.36 | 6.51 | 3,012 | a3f2c08a | BEST_VS_BEST_GPU_20261112.md:19-21 |
| 0x13BC + M1361 | W3 d11 TS1 | 782.20 / 787.55 | 7.18 | 2,724 | tracker :1681 | |
| 0x13BC + M1361 | K_C 128 / 256 / 256 TS0 (W2 d11) | 903.45 / 827.13 / 760.59 | 6.88 / 6.87 / 7.00 | 2,851 / 2,857 / 2,803 | 1b0779cf / 1b1e44c1 / a2cc35a1 | accept_gap_20261112/silicon/ag_table.txt |
| 0x13BC + M1361 | W1 d11 V12 / d15 V16 TS1 | 719.34 / 720.03 | 5.59 / 5.83 | 3,506 / 3,365 | baf31f4e | DESIGN_FABLE_ACCEPT_NOLEVEL_20261112.md S2 |
| GPU-only (same H200) | d9 / d11 / d13 fa3 | 841.27 / 825.91 / 778.82 | 6.63 / 7.05 / 7.26 | 2,960 / 2,776 / 2,696 | BEST_VS_BEST_GPU_20261112.md; F3T_CONFIG_SWEEP_20261109.md:122 |
- Texts and steps are identical across ATT images at the same MLPV and config: COMBO3 / D5F / 0x13BE all give 1b0779cf at W2 d11 and 54f88d49 at W4 TS0. So accept and steps carry over from 0x13BE / 0x13BC cells; tok/s does not.
- Has W3 run on 7 cards? Yes, with TREE_STOP on: 0x13BC 11-13, 0x13BE 11-14, and SIBW B3 11-13 (728.79 serial). W3 with TREE_STOP off: never. W3 on COMBO3: never.
- W > 4: W8 ran (0x13BE, and C1 W8 427.4). W5-7 never ran. The RTL and host allow W up to 8 with ALLOW_WIDE.
- Never measured on COMBO3: W2 d7 / d9 / d13, W3 at any depth, W4 at d != 11, W1 at any depth, every K_C != 64.
3. Kit (create-only, kcdw_20261116/ship/)
| file | md5 | what |
|---|---|---|
derive_kcdw.py |
1effaa1f | count-asserted derivation from the frozen window scripts (asserts ac_serve_w.sh 07602f90, ac_steps.sh 4b69ec38 = the talon2 att_combo_20261116/ copies, MD5SUMS -c clean 13:29Z) |
kcdw_serve_w.sh |
b5f63650 | = ac_serve_w.sh + per-round fields, out dir $RUN/serve_kcdw, SUMMARY cols 16-20, post-round KC_CHECK (m1_kc_cfg EQUAL on 50/63 AND == round K_C, else NO_REARM_kc_* and stop). Diff kcdw_serve_w.diff (14cbe805) |
kcdw_steps.sh |
abebc903 | = ac_steps.sh; serve / smoke call kcdw_serve_w.sh; NEW phase kcgolden <kc> (record golden at that K_C; the 1st of the boot is that K_C's reference, the 2nd must match it; exact tooth; M1 read; NO_REARM on hang). Diff kcdw_steps.diff (e6b9bfb5) |
kcdw_table.py |
fc1b5092 | per-cell table (W, d, K_C, TS, theta, V) + anchor sequence; failed rows never averaged |
kcdw_mk_head512.sh |
7ccaae5c | config-only K_C 512 clone of the card-head twin ckpt/head256 (978541fb) for the GPU-only arm |
Zero-install path: the frozen ac_steps.sh serve ALREADY takes K_C as ROUNDS field 6 (ac_serve_w.sh:37,49). These run today with no new file, at W2, TS1 and theta -2.5 only:
bash $T2ROOT "STAGE=d6 RUNB=<run base> ROUNDS='a5:11:fa3:4:1:64 ka2:11:fa3:4:1:128 kb2:11:fa3:4:1:256 a6:11:fa3:4:1:64 k512n:11:fa3:4:1:512 a7:11:fa3:4:1:64' bash $TW/ac_steps.sh serve"- Use the 512 row only after a clean
kcgolden 512, or accept the risk and place it last.
The kcdw files are needed for: theta re-tune, TS0, width, VTOK per round, KC_CHECK and kcgolden.
ROUNDS syntax (kcdw): tag:depth:backend:fastglue:cudapack:kc:ts:width:vtok:theta.
- Defaults: kc 64, ts 1, width $SW (2), vtok 20, theta -2.5. Old 5-field strings behave as before.
- Width > 2 adds
EAGLE4_4C1_ALLOW_WIDE=1per round. bash -nis clean on both scripts. The field parser was unit-run on 4 sample rounds. Not yet run on talon2.
Install (box1, create-only, before the window; T2 / T2ROOT / TW as in WINDOW_RUNBOOK_ATTCOMBO_20261116.md:39):
for f in kcdw_serve_w.sh kcdw_steps.sh kcdw_table.py kcdw_mk_head512.sh; do
$T2 "f=$TW/$f; [ -e \$f ] && { echo EXISTS \$f; exit 3; }; cat > \$f; md5sum \$f" < reports/results/DRAFTER_SPEED_RESEARCH_20261107/kcdw_20261116/ship/$f
done
E="STAGE=d6 RUNB=<this boot's run base, e.g. runs/acwin_20261116b> KTAG=kcdw" (the keeper boot's env, ATT_COMBO_SILICON_20261116.md last line).
4. Window plan (cells ordered safe first, untested last; HANG RULE)
HANG RULE (coordinator 11-13): any rc != 0, errs != 0, KVRD_ABORT or KC_CHECK FAIL writes NO_REARM. The serve loop stops, and every later arm / golden / serve phase refuses on this boot.
- Next step is a cold AC, then runbook steps 5-8 (drv qdma status2 ident, kvread, bringup with fp 51524c4f + tooth, d15 smoke).
- No re-flash is needed: COMBO3 is the keeper.
- Resume at the cell after the one that failed. Never re-arm onto a hung card.
Before B, if this boot has no T4_GOLDEN_OK: runbook steps 5-8. Then read 0x15C on 50/63 (section 1). Part A below is box1-two-v2's running grid. Do not re-run it. Its rows are listed only as kcdw ROUNDS for any cell the grid skipped. Pick C1 / C2 / C3 from the grid's table (DW_SWEEP_7CARD_COMBO3_20261116.md).
Rounds: 2 per new cell, the pass reversed on the second round (ABBA against drift), and a W2 d11 TS1 K_C 64 anchor aN between phases.
- W1 uses VTOK = d+1. Everything else uses V20.
- TREE_STOP per the measured rule: TS1 at W <= 2, TS0 at W >= 3 (W2: TS0 887 vs TS1 936 = -5.2 %; W4: TS0 756.05 vs TS1 714.59 = +5.8 %,
W4_SERVE_BOUND_20261115.md:210-227). - W1 also takes the host verdict (loop_seq W != 2), so A4 adds a W1 TS0 probe.
| phase | cells (new, x2 unless noted) | risk class | command (bash $T2ROOT "$E ROUNDS='...' bash $TW/kcdw_steps.sh serve") |
|---|---|---|---|
| A1a (grid) | anchor; W2 d9, d13, d7 TS1; W2 d15 TS1 (x1, plus the smoke = 2); W4 d9, d13, d7, d15 TS0 | paths run today on COMBO3 (W2 TS1 card stop at d11 / d15; W4 TS0 at d11) | a0:11:fa3:4:1:64:1:2:20:-2.5 w2d9a:9:fa3:4:1:64:1:2:20:-2.5 w2d13a:13:fa3:4:1:64:1:2:20:-2.5 w2d7a:7:fa3:4:1:64:1:2:20:-2.5 w2d15b:15:fa3:4:1:64:1:2:20:-2.5 w4d9a:9:fa3:4:1:64:0:4:20:-2.5 w4d13a:13:fa3:4:1:64:0:4:20:-2.5 w4d7a:7:fa3:4:1:64:0:4:20:-2.5 w4d15a:15:fa3:4:1:64:0:4:20:-2.5 a1:11:fa3:4:1:64:1:2:20:-2.5 |
| A1b (grid) | second pass, reversed | same | w4d15b:15:fa3:4:1:64:0:4:20:-2.5 w4d7b:7:fa3:4:1:64:0:4:20:-2.5 w4d13b:13:fa3:4:1:64:0:4:20:-2.5 w4d9b:9:fa3:4:1:64:0:4:20:-2.5 w2d7b:7:fa3:4:1:64:1:2:20:-2.5 w2d13b:13:fa3:4:1:64:1:2:20:-2.5 w2d9b:9:fa3:4:1:64:1:2:20:-2.5 a2:11:fa3:4:1:64:1:2:20:-2.5 |
| A2 (grid) | W3 d11, d9, d13, d7, d15 TS0 | W3 ran TS1 on 7 cards; W3 TS0 and W3 on COMBO3 never | w3d11a:11:fa3:4:1:64:0:3:20:-2.5 w3d9a:9:fa3:4:1:64:0:3:20:-2.5 w3d13a:13:fa3:4:1:64:0:3:20:-2.5 w3d7a:7:fa3:4:1:64:0:3:20:-2.5 w3d15a:15:fa3:4:1:64:0:3:20:-2.5 w3d15b:15:fa3:4:1:64:0:3:20:-2.5 w3d7b:7:fa3:4:1:64:0:3:20:-2.5 w3d13b:13:fa3:4:1:64:0:3:20:-2.5 w3d9b:9:fa3:4:1:64:0:3:20:-2.5 w3d11b:11:fa3:4:1:64:0:3:20:-2.5 a3:11:fa3:4:1:64:1:2:20:-2.5 |
| A3 (grid) | W1 d11 V12, d15 V16, d9 V10, d13 V14, d7 V8, TS1 | W1 ran at d11 V12 / d15 V16 (0x13BC); V8 / V10 / V14 never (a wrong V = host rc 247, card clean) | w1d11a:11:fa3:4:1:64:1:1:12:-2.5 w1d15a:15:fa3:4:1:64:1:1:16:-2.5 w1d9a:9:fa3:4:1:64:1:1:10:-2.5 w1d13a:13:fa3:4:1:64:1:1:14:-2.5 w1d7a:7:fa3:4:1:64:1:1:8:-2.5 w1d7b:7:fa3:4:1:64:1:1:8:-2.5 w1d13b:13:fa3:4:1:64:1:1:14:-2.5 w1d9b:9:fa3:4:1:64:1:1:10:-2.5 w1d15b:15:fa3:4:1:64:1:1:16:-2.5 w1d11b:11:fa3:4:1:64:1:1:12:-2.5 a4:11:fa3:4:1:64:1:2:20:-2.5 |
| A4 (grid, optional) | W1 d11 TS0 | W1 TS0 never | w1d11z1:11:fa3:4:1:64:0:1:12:-2.5 w1d11z2:11:fa3:4:1:64:0:1:12:-2.5 |
| read the grid | the DW_SWEEP table (or kcdw_table.py <RUN>/serve_kcdw for kcdw rows) |
choose B cells: C1 = W2 d11 TS1 (fixed: it has cross-image K_C references), C2 = the fastest W2 depth != 11, C3 = the fastest W>=3 TS0 cell (the stop-free accept read of K_C) | |
| B0 (GPU only, no card) | twin captures + K_C 512 twin: g_head64 / g_head256 d11 with CAP=, g_head512 d11, g_head256 d<C2> |
none (H200 only; run inside the card lock as on 11-12) | $T2 "bash $TW/kcdw_mk_head512.sh", then accept_gap_20261112/ship/ag_gpu_cell.sh with FAST=~/eagle4_talon2_4c1/accept_gap_20261112/ckpt/head{64,256,512} CAP=<prefix> (the 11-12 GPU-arm invocation, ag_gpu_cells.sh) |
| B1 | kcgolden 128 x2, kcgolden 256 x2 (separate calls) |
K_C 128 / 256 ran clean on M1361 | bash $T2ROOT "$E bash $TW/kcdw_steps.sh kcgolden 128" (repeat; then 256): KC128 REPEAT ... PASS = deterministic; tooth RC 0 |
| B2 C1 | cross-image checks at theta -2.5 (x1), then re-tuned theta (x2), alternating with K_C 64 anchors | ran clean on M1361 | a5:11:fa3:4:1:64:1:2:20:-2.5 ka2:11:fa3:4:1:128:1:2:20:-2.5 kb2:11:fa3:4:1:256:1:2:20:-2.5 a6:11:fa3:4:1:64:1:2:20:-2.5 k128r1:11:fa3:4:1:128:1:2:20:-2.58 k256r1:11:fa3:4:1:256:1:2:20:-2.65 a7:11:fa3:4:1:64:1:2:20:-2.5 k256r2:11:fa3:4:1:256:1:2:20:-2.65 k128r2:11:fa3:4:1:128:1:2:20:-2.58 a8:11:fa3:4:1:64:1:2:20:-2.5 |
| B3 C3 | K_C 128 / 256 at TS0 (no theta interaction = the clean accept number) | ran clean (kz1 = 256 TS0) | kz128a:<d>:fa3:4:1:128:0:<W>:20:-2.5 kz256a:<d>:fa3:4:1:256:0:<W>:20:-2.5 kz64m:<d>:fa3:4:1:64:0:<W>:20:-2.5 kz256b:<d>:fa3:4:1:256:0:<W>:20:-2.5 kz128b:<d>:fa3:4:1:128:0:<W>:20:-2.5 |
| B4 C2 | K_C 128 / 256 re-tuned at the best W2 depth | new depth, known K_C | c2a:<dB>:fa3:4:1:64:1:2:20:-2.5 c2k128a:<dB>:fa3:4:1:128:1:2:20:-2.58 c2k256a:<dB>:fa3:4:1:256:1:2:20:-2.65 c2k256b:<dB>:fa3:4:1:256:1:2:20:-2.65 c2k128b:<dB>:fa3:4:1:128:1:2:20:-2.58 c2b:<dB>:fa3:4:1:64:1:2:20:-2.5 |
| B5 (LAST) | kcgolden 512 x2 (= the K_C 512 smoke: the first K_C 512 trees ever on G8_OWN silicon, 74 score beats per item), then C1 512 re-tuned x2 + theta -2.5 x1, C3 512 TS0 x2, C2 512 re-tuned x2 |
never on silicon | kcgolden 512 (x2) then k512r1:11:fa3:4:1:512:1:2:20:-2.72 a9:11:fa3:4:1:64:1:2:20:-2.5 k512n:11:fa3:4:1:512:1:2:20:-2.5 k512r2:11:fa3:4:1:512:1:2:20:-2.72 kz512a:<d>:fa3:4:1:512:0:<W>:20:-2.5 kz512b:<d>:fa3:4:1:512:0:<W>:20:-2.5 c2k512a:<dB>:fa3:4:1:512:1:2:20:-2.72 c2k512b:<dB>:fa3:4:1:512:1:2:20:-2.72 a10:11:fa3:4:1:64:1:2:20:-2.5 |
Round count for the K_C part: B 28 serves + 6 goldens + 4 GPU cells, about 45 min warm, 1 h 15 min budget (D9). If the window is short: run B1 + B2 (the C1 cross-image checks) + B5 kcgolden 512 first, and cut B4.
5. Correctness per K_C (what replaces golden_fp_gate)
golden_fp_gate.pyis the K_C 64 record reference. It hashes K_C, so K_C != 64 has no built-in reference (golden_fp_gate.py:77-84; K_C 32 hashed de4e5697 by design).- What stands in its place:
- The image gate stays the K_C 64 bring-up gate: fp 51524c4f + EXACT_TOOTH x2 on this boot. K_C is a runtime bound in the same datapath, xsim MATCH at 128 / 256 / 512.
kcgolden <kc>x2 per K_C. The first run is the reference of that K_C on this boot; the second must hash identical (deterministic, 0 discarded, gate rc 0). EXACT_TOOTH must pass; it is K_C-robust: T1 dominant token at rank 0, T2 L0 mass within 0.5 log2, and the K_C 32 control moved the mass by <= 0.10.- Cross-image serve references (same MLPV, same config, ATT-independent texts):
ka2must give 2,851 steps and texts 1b0779cf;kb2must give 2,857 steps and texts 1b1e44c1 (the 11-12 rows). TS0 K_C 256 at W2 d11 would give 2,803 / a2cc35a1. - Every round: KC_CHECK OK (0x150 = K_C on both MLPVs), 0x1D4 0x3000 x4, rc 0, errs 0.
- K_C 512 has no prior reference: the two kcgolden 512 must match each other, and the two rounds of each 512 cell must be text- and step-identical.
6. If the user wants accept from K_C to pay: levers with the right shape [D, not built]
- The cost is G8 exposure: about 24 cycles per candidate per item past K_C ~128 (D2). Each lever has to cut that slope:
- (a) More gather channels or a wider owner split (G8_NCH 4 -> 8, or more lm-head owners). This halves the per-candidate cycles.
- (b) Per-level K_C: a large K_C only at the levels where the shortlist binds. The card already samples kcq per item at G8 launch (
g8_rescore_p.sv:1604), so this needs a per-depth table register, still written equal on both MLPVs. - (c) A cheap head that makes K_C 64 lossless: finer G12 levels or grouped scales. This targets the 18-step card-head term directly, at no G8 cost.
- B0 decides between them. If the twin's K_C 256 / 512 captures pick children the card does not, the defect is upstream of G8 (prefilters or the bound), and (c) or a prefilter fix is the lever. If they agree, the 15 vs 3 discrepancy sits elsewhere.
- K_C > 512 needs a rebuild:
EAGLE4_G8_K_C 1024, sel / rescored / FG8 key_mem x2, and an 11-bit kcq on the own link. It is not recommended: D7 already prices 512 at -23 %.
Evidence index
- RTL: the staged M1361 tree above (md5s in section 1). Stage record
/mnt/models/stage_E4B71361/STAGE_SHA_E4B71361.txt. - Prior silicon:
accept_gap_20261112/silicon/{ag_table.txt,serve_ag/SUMMARY.tsv (26bb71b9),serve_ag/SUMX.tsv,gpu/GPU_POINTS.tsv},DESIGN_FABLE_ACCEPT_GAP_7C_20261112.md:240-275,DESIGN_FABLE_ACCEPT_NOLEVEL_20261112.md(S1-S3),KC32_7CARD_AB_20261111.md,W4_SERVE_BOUND_20261115.md:197-231,ATT_COMBO_SILICON_20261116.md,FOURCARD_DESIGN_20261016/4C1_DELIVERY_TRACKER.md:1657,1671-1681,1705-1707,1735. - Captures for B0:
accept_gap_20261112/silicon/serve_ag/cap_cq1.8171.jsonl,cap_cp1.9205.jsonl(+ box1/mnt/data/eagle4_scratch/nolevel_win_20261112/), toolaccept_nolevel_20261112/kc_cap_compare.py(6c1eca3b); twin ckpts box1/mnt/data/eagle4_scratch/accept_gap_split_20261019/ckpt/head256(978541fb, num_candidates 256), talon2~/eagle4_talon2_4c1/accept_gap_20261112/ckpt/. - Window kit parents:
attcombo_20261116/ship/(MD5SUMS.txt 549d83fc).