# 7-card depth x width (+ K_C) sweep on COMBO3 0xE4B713D6 + M1361, 2026-11-16

Program 1: the 7-card Llama-3.2-3B drafter on talon2. The user ordered this sweep: "Also need a sweep on depth and width for 7 card".
- Cards: ATT COMBO3 0xE4B713D6 d2e737d7 x4 on 17/3d/bd/e1, plus MLPV M1361 0xE4B71361 e24deb3f x2 on 50/63. This set is RUNNING and STORED.
- Boot **321f3241** (the keeper boot of the ATT-combo window): IDENT 6/6, **GOLDEN_FP PASS 51524c4f**, EXACT_TOOTH x2.
- The W2 d11 rows on boot 79cc9d1b (t1-t3, off1, s15, w4a/b) are from the same image, flashed earlier the same day (`ATT_COMBO_SILICON_20261116.md`).

Bench: humaneval 80x256, fa3, fastglue 4, cudapack 1, frozen kit `att_combo_20261116/kit_D6`, record levers HX_MMIO7/O7/S7 v, KV1H2C on, theta -2.5.
- TREE_STOP: on for W1/W2, off for W3/W4 (TS=0 via `ac_steps_ts.sh` / `ac_serve_w_ts.sh`, which differ only in `TREE_STOP=${TSTOP:-1}`).
- VTOK: 20, except W1 = depth+1.
- W>2 uses EAGLE4_4C1_ALLOW_WIDE=1.
- Every round: rc 0 / errs 0; ATT 0x1D4 0x00003000 x4 after the round (KV1_VERDICT CLEAN); per-request texts md5 (80); MLPV 0x150 == the requested K_C on both MLPVs. **54 rows, 0 aborts, 0 hangs, 0 NO_REARM.**

Raw rows: `dw_sweep_combo3_20261116/DW_SWEEP_ROWS_COMBO3_20261116.tsv` (md5 104e5d73; talon2 copy `runs/acwin_20261116b/dw_sweep/`, built by `dw_rows.py` from both serve dirs' SUMMARY/SUMX). Per-cell talon2 logs: `runs/acwin_20261116b/dw_sweep/logs/`.

## tok/s (mean of rounds), K_C 64
| W \ depth | 7 | 9 | 11 | 13 | 15 |
|---|---|---|---|---|---|
| **W1** TS1 | 698.86 @4.90 | 716.34 @5.31 | 725.07 @5.59 | **728.12** @5.75 | 726.88 @5.83 |
| **W2** TS1 | 888.03 @5.91 | 932.01 @6.51 | **960.19** @6.91 (n=6) | 960.00 @7.07 | 958.68 @7.17 |
| **W3** TS0 | 868.91 @6.18 | **883.63** @6.80 | 868.43 @7.20 | 834.62 @7.38 | 790.81 @7.46 |
| **W4** TS0 | **814.90** @6.24 | 813.48 @6.84 | 792.50 @7.20 | 753.49 @7.41 | 709.96 @7.50 |

- n = 2 per cell, except W2 d11 n = 6 (960.23 / 960.97 / 963.10 / 956.62 / 957.42 / 962.78, across 2 boots).
- Round-to-round spread is ≤ 5 tok/s in every cell.
- Texts are identical across the rounds of every cell. They differ between configurations: W1 baf31f4e at all depths; W2 d11 1b0779cf (= record); W4 d11 54f88d49 (= record W4).
- **The best cell is W2 d11-d13 (960)**; W2 d15 is within 0.2 %.
- W1 is flat from d11 on (725-728). W3 peaks at d9 and W4 at d7-d9. Deeper trees raise accept but lose tok/s at W ≥ 3.
- **W4 d11, TREE_STOP on vs off:** TS1 721.25 @7.19 (1 round), TS0 792.50 @7.20, so **TS0 is +9.9 %**. Texts 54f88d49 both ways.
- **KV1H2C off** (W2 d11, 1 round): 955.22 vs 960.19 on.

## K_C cells (field 6 of ROUNDS; MLPV 0x150 read back = K_C on both MLPVs every round)
| cell | K_C 64 | K_C 128 | K_C 256 | K_C 512 |
|---|---|---|---|---|
| W2 d11 TS1 | 960.19 @6.91, 2839 steps, 1b0779cf | **927.38 @6.88** (2), 2851, 1b0779cf | **843.89 @6.87** (2), 2857, 1b1e44c1 | **708.79 @6.86** (1), 2860, 1b0779cf |
| W3 d9 TS0 | 883.63 @6.80, 2884 | 830.69 @6.81 (1), 2886, 8c602a11 | 723.61 @6.80 (1), 2888, 8c602a11 | not run |

- The W2 d11 K_C 64 rounds bracket the K_C 128/256 rounds (957.42 before, 962.78 after), so there is no drift.
- **K_C 512 ran on silicon for the first time: clean** (rc 0, errs 0, KV CLEAN, 0x150 = 0x200 x2).
- Golden fp 51524c4f covers K_C 64 only. The other K_C values have texts-md5 consistency only.
- The 11-12 expectations (0x13BC + M1361) were 903 / 827. COMBO3 measures 927.38 / 843.89.
- MLPV 0x15C, read-only c4 status before the K_C 512 round: **0xacf10000 on MLPV0 and MLPV1 (bit2 = 0)**. ATT0 reads the same. File `runs/acwin_20261116b/dw_sweep/status_15c_pre512.json`.

## Census + verify timing (COMBO3 + M1361, boot 321f3241)
- **Per-phase census** (read-only bench `c4.py bench --md 11 --width 2 --kc 64 --trees 100 --census 20`, record modes TS=0; script `att_combo_20261116/ac_census_combo3.sh` 961cfd42):
  - JSON `runs/acwin_20261116b/census_combo3/bench_P11w2_md11_w2.json` (**md5 a0603abba0ebbce88909a88f7402fca0**). Table: `dw_sweep_combo3_20261116/census_combo3_P11w2.table.txt` (7a878794).
  - 100 trees ok, 0 discarded, 0 nondeterministic. issue_done p50 **2.622 ms/tree** (381.1 trees/s), 238 us/level, 125 us/item.
  - ATT0 us/tree: G1 145.09 · G4 452.14 · G5 451.89 · G6 220.48 · G7g 332.18 · G7u 177.23 · G7p 42.32 · PIPE_IO 192.48.
  - MLPV0 us/tree: G9 270.45 · G11 148.27 · GLM 71.91 · G12 7.00 · TREE 707.64 · G8 942.09 (G8_TAIL = G8 − TREE 234.45) · PIPE_IO 169.52.
- **GPU verify time** (`EAGLE4_VERIFY_TIMING=1` via the kit's chaining shim `scripts/4c1/sglang_site_bv`, one W2 d11 round):
  - **961.06 @6.91**, texts 1b0779cf, so the timer does not slow the serve.
  - `runs/acwin_20261116b/dw_sweep/vtiming/verify_ms.csv` (md5 04455237, 2,876 rows): verify-only **p50 3.451 ms/step** (p10 3.441, p90 3.463).
  - Sum 9.97 s over a 20.41 s run with 2,839 steps (= 7.19 ms/step total).

## Calculations and hypothesis
- [C] Verify takes 3.45 of the 7.19 ms per step (48 %). The rest (≈ 3.7 ms/step) is the drafter tree path + host.
- [H] W ≥ 3 loses at depth because the per-level cost grows with width while accept saturates (accept 7.20 → 7.50 from d11 to d15 at W4, while tok/s falls 10 %). Not measured per level here.
