Independent re-verification (2026-10-12) + fixes

Re-ran the full 4/8/16/32 GiB memory matrix on the same host, same
llama.cpp commit 836d5717, same patch, same reserve (SDQ_RESERVE_MB=2048),
built with the portable ISA flags from section 4.2. All 8 runs clean; every
published column reproduces within noise. arena / slots / peak RSS / total
SSD are near bit-identical; tok/s within +-6%.

Fixes and corrections:
* sdq-dsv4/verify_dequant_dsv4.py: TARGETS was one flat suffix list applied
to BOTH layer 0 and 26; ffn_down_exps is MXFP4 in both, so 3 comparisons
ran twice -> printed 21 rows for 15 unique comparisons and self-reported
"21/21". README said "already fixed" but only the results file had been
edited. Now per-layer explicit, with two guards; prints 15/15.
* sdq-dsv4/determinism_probe.sh: hardcoded non-existent /root/app/work
paths; now auto-detects and fails loudly.
* README 5(e): the claim that "AMX" does not appear in libggml-cpu.so is
false (AMX_INT8 comes from the CPUID probe in cpu-feats.cpp). Conclusion
unchanged; evidence replaced with nm -D on registered symbols.
* README 6.3: 387/387 is prompt-specific, not a general property. With
"The capital of France is" only prefill is bit-identical (43/387).
Section A itself reproduces exactly and was not wrong.
* README 6.4: --ram-budget-mb belongs on the list of knobs that perturb
decode output; it was missing. Compile flags excluded as the cause
(two builds disagree too).
* README section 3: the build caveat is resolved -- the matrix does hold
for the section 4.2 portable flags.
* README 4.1: `hf download` aborts on typer>=0.27 + click<8.2; documented
the Python-API fallback. Unrelated to the token.
* README "rerun" section: item (3) was labelled as the MoE A/B but ran the
determinism probe; relabelled and the MoE A/B added as item (4).

Original published results files are left untouched.

README.md

@@ -112,6 +112,24 @@ decode 期不確定性的一部分，不同 RAM 預算／無關設定會換掉
 > 這個矩陣另外用了 `SDQ_RESERVE_MB=2048`，而 pager 的**程式預設值是 900 MB**
 > （`sdq_pager.cpp`）。這個值直接決定 arena 大小（`arena = 預算 − RSS − 預留`），
 > 不帶這個環境變數重跑會得到比表中更大的 arena 與不同的槽位數。
+> ✅ **這個保留意見可以撤掉了（2026-10-12 複驗）**：用第 4.2 節那組**可攜 ISA
+> 選項**（`-DGGML_NATIVE=OFF -DGGML_AVX512=ON ...`）重跑整個 4/8/16/32 GiB
+> 矩陣，8 輪全部無錯，**每一欄都在雜訊內重現**（發布值→複驗值）：
+>
+> | RAM 預算 | arena (GiB) | slots | decode tok/s | hot % | SSD MiB/tok | peak RSS (MB) |
+> |---:|---|---|---|---|---|---|
+> | 4 GiB | 1.79→1.79 | 164→164 | 0.69→0.65 | 0.0→0.0 | 2944.4→2944.4 | 2233→2232 |
+> | 8 GiB | 5.79→5.79 | 531→531 | 1.40→1.41 | 42.3→42.4 | 1451.6→1444.7 | 6325→6325 |
+> | 16 GiB | 13.79→13.79 | 1264→1264 | 1.63→1.62 | 54.5→54.9 | 1021.8→1007.5 | 14508→14507 |
+> | 32 GiB | 29.79→29.79 | 2732→2732 | 1.71→1.72 | 63.9→64.0 | 690.7→688.0 | 30894→30894 |
+>
+> 決定性的欄位（arena、slots、peak RSS、SSD 總量）幾乎是位元級相同；只有 tok/s
+> 落在 ±6% 內，這與 4.2 節自己承認的「同一組設定重跑三次就差 1.66×」是同一個
+> 量級的抖動。**所以這張表對第 4.2 節的編譯選項同樣有效。**
+>
+> 順帶釐清一處 JSON 內部矛盾：`build.cmake` 欄位寫的是第 4.2 節那組可攜選項，
+> `build.note` 卻說是用 `-march=native -mno-amx-*` 量的。複驗顯示兩者都對得上這張
+> 表，矛盾不影響數字，但 `build.note` 那句「tok/s 要照實說明」已無必要。
 
 > 三個容易搞混的地方，這裡先講清楚：
 > 1. `nproc` 在這台機器上回 16，是因為環境變數 `OMP_NUM_THREADS=16`，
@@ -166,6 +184,35 @@ hf download unsloth/DeepSeek-V4-Flash-0731-GGUF --include "UD-IQ4_NL/*" \
    --local-dir ./models
 ```
 
+> ⚠️ **`hf` CLI 在某些環境會直接掛掉，要改用 Python API**（2026-10-12 實測）。
+> 症狀是在開始下載之前就丟出：
+>
+> ```
+> File ".../typer/core.py", line 725, in parse_args
+>     ctx._opt_prefixes.update(parser._opt_prefixes)
+> AttributeError: 'StyledContext' object has no attribute '_opt_prefixes'
+> ```
+>
+> 這是 `typer` 與 `click` 的版本不相容（typer ≥ 0.27 需要 click ≥ 8.2 的
+> `Context._opt_prefixes`，但環境裡是 click 8.0.4），**跟 HF token、
+> 權重、網路都無關**，改 token 沒用。權重是 public repo，其實根本不需要 token。
+> 兩種解法：
+>
+> ```bash
+> # (a) 直接用 Python API（不經過 typer，最省事）
+> python3 -c "
+> from huggingface_hub import snapshot_download
+> print(snapshot_download('unsloth/DeepSeek-V4-Flash-0731-GGUF',
+>                        allow_patterns=['UD-IQ4_NL/*'], local_dir='./models'))"
+>
+> # (b) 升級 click
+> pip install -U 'click>=8.2'
+> ```
+>
+> 下載完記得核對 4 個 shard 的 bytes 應為
+> `5257664 + 49431060672 + 49605340544 + 37620787776 = 136662446656`，
+> 對不上就是下載沒完（本 repo 實測下載速率約 1.5 GB/s，約 92 秒完成）。
+
 ### 4.2 建引擎
 
 ```bash
@@ -347,7 +394,23 @@ ggml_backend_sched_split_graph: GGML_ASSERT(*cur_backend_id != -1) failed
 
 * **推薦（4.2 採用）**：不要用 `-march=native`，改成明確列 ISA
   （`-DGGML_NATIVE=OFF -DGGML_AVX512=ON ...`）。AMX 的三個 CMake 選項本來就
-  預設 OFF，`libggml-cpu.so` 裡連 `"AMX"` 這個字串都不會出現，自然不會註冊。
+  預設 OFF，所以 AMX extra buffer type 不會被註冊。
+
+  > ⚠️ **勘誤（2026-10-12 複驗）**：本節原本寫「`libggml-cpu.so` 裡連 `"AMX"`
+  > 這個字串都不會出現」——**那句是錯的**，該字串確實會出現。
+  > `libggml-cpu.so` 裡有一個 `AMX_INT8` 字串，來自 ggml 的 CPUID 機型偵測
+  > （`ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp` 的 `bool AMX_INT8(void)`），
+  > 那段程式碼本來就**一定**會編譯進來，與 AMX buffer type 有沒有註冊無關。
+  > **正確的判準是看有沒有註冊的符號**，本機實測：
+  >
+  > ```bash
+  > $ nm -D --defined-only build/bin/libggml-cpu.so | grep -i amx
+  >   000000000001d400 T ggml_cpu_has_amx_int8     ← 只有這個 CPUID 探測
+  > $ strings build/bin/libggml-cpu.so | grep -c 'amx_extra_buffer_type'
+  >   0                                        ← AMX buffer type 未註冊
+  > ```
+  >
+  > 結論（AMX 必須關）不變，錯的只是拿來佐證的那句話。
 * 如果你一定要 `-march=native`：加
   `-DCMAKE_C_FLAGS="-mno-amx-int8 -mno-amx-tile -mno-amx-bf16"`（CXX 同樣）。
 
@@ -403,7 +466,16 @@ python3 sdq-dsv4/verify_dequant_dsv4.py
 > 這裡是 5 個張量 × 3 個 row = **15 列**。原始輸出檔
 > `results/verify_dequant.txt` 早先印成 21 列，但其中 `blk.0.ffn_down_exps` 與
 > `blk.26.ffn_down_exps` 各被重��印了一次（每個 3 列，共 6 列重複），
-> 所以「21/21」是重複計數造成的，不是真的做了 21 次獨立比對。已修正。
+> 所以「21/21」是重複計數造成的，不是真的做了 21 次獨立比對。
+>
+> ✅ **2026-10-12：根因已修掉（先前只改了結果檔，腳本沒動）**。
+> 根因是 `verify_dequant_dsv4.py` 把 `TARGETS` 寫成一張「張量後綴」清單，
+> 然後外層 `for il in (0, 26)` 把整張清單套用到**兩層**；而 layer 0 與 layer 26
+> 的 `ffn_down_exps` 都是 MXFP4(39)，所以那 3 列被比對了兩次 → 印 21 列、實際 15 列。
+> 已改成 `(層, 後綴, 型別, 元素數)` 逐項寫死，並加上兩道防呆：
+> 張量不存在或型別不符會直接 `SystemExit`（原本是 `continue` 靜靜跳過），
+> 以及比對結束後檢查「列數 == 唯一張量數」，重複就直接報錯。
+> 修完重跑輸出 **15/15 bit-exact**，與本節表格與 `results/verify_dequant.txt` 一致。
 
 ### 6.2 整個 expert 前向對照
 
@@ -451,6 +523,30 @@ hot/cold 比例、SSD 讀取量、槽位數（164 vs 2732）、淘汰次數全
 改變數學」最直接的證據。逐檔 md5 見
 [`results/moe_dump_bitexact.md`](results/moe_dump_bitexact.md)。
 
+> ⚠️ **重要限定：387/387 是「這個 prompt」才有的結果，不是通則**（2026-10-12 複驗）。
+> 把 §A 的設定整個照抄、只換 prompt 成 `The capital of France is`，
+> 得到的**不是** 387/387，而是：
+>
+> | prompt | prefill bit-identical | decode bit-identical | 合計 |
+> |---|---:|---:|---:|
+> | `Explain mixture-of-experts routing in one sentence.`（§A 原文） | 43 / 43 | **344 / 344** | **387 / 387** |
+> | `The capital of France is` | 43 / 43 | **0 / 344** | 43 / 387 |
+>
+> 兩邊都是 `--ram-budget-mb 4096` vs `32768`、其餘設定完全相同。
+> 第一列**精確重現**了 §A 的 387/387，所以 §A 本身沒有問題。
+> 第二列說明：換成 top-1/top-2 本來就幾乎並列的 prompt，decode 期就會分岔。
+> 分岔點同樣在**第一個 decode call** 就出現（`d.0.43`，也就是 layer 0 的第 1 次
+> decode），此時 `cur` 的 `max|Δ|` 相對值已經是 2.1（不是微小浮點雜訊），
+> `ids` 也跟著換成別組 expert。
+>
+> 合理的解釋是：prefill 結束後的 logits 有極小差異 → `--temp 0` 的 argmax 換成
+> 另一個同義 token → 後續 decode 的輸入 `cur` 整條岔開。這與 6.4 節的結論一致
+> （分歧來自 llama.cpp 的 decode 期 dsv4 CSA/HCA 路徑，不是 SDQ 的分頁）。
+>
+> 所以這節標題「分頁結果與 arena 大小無關」要這樣讀：
+> **對這個 prompt 成立；對 top-1/top-2 接近的 prompt 不成立。**
+> 引用 387/387 這個數字時必須連 prompt 一起引用，否則會被誤讀成端到端通則。
+
 另外同一設定重跑 6 次、兩個 RAM 預算各 6 次，greedy（`--temp 0`）輸出
 **12/12 逐字相同**（`results/determinism.txt`）。
 
@@ -466,6 +562,29 @@ hot/cold 比例、SSD 讀取量、槽位數（164 vs 2732）、淘汰次數全
 `SDQ_IO_SPLIT=2`、`-c`、`-b`），第一個 token 就會換。`-t 1`（完全單執行緒）
 也一樣會換，所以不是執行緒競爭。
 
+> ⚠️ **`--ram-budget-mb` 也要加進這份清單（2026-10-12 複驗新增）**。
+> 本來這份清單裡沒有 RAM 預算，本 repo 的 `results/determinism.txt` 也顯示
+> 8192 MB 與 32768 MB 兩個預算**輸出相同**（同為 `0e637837ac89`）。
+> 但在本機用 4.2 的可攜編譯選項重跑，每個預算**各自**仍是 6/6 逐字相同，
+> **兩個預算彼此卻不同**：
+>
+> | build | 8192 MB | 32768 MB | 是否相同 |
+> |---|---|---|---|
+> | 可攜 ISA（第 4.2 節） | `08c061788823` | `175b890dddeb` | **不同** |
+> | `-march=native -mno-amx-*` | `d5cf2396b424` | `54a418d53304` | **不同** |
+>
+> 兩種 build 都一樣，所以**不是編譯選項造成的**（原本懷疑是，已排除）。
+> 實際內容：8192 MB 給出 `capital of France is **Paris**.`，
+> 32768 MB 給出 `'m sorry, I can't answer that question.`。
+> 配合 6.3 節新加的限定表，這是同一件事的兩種表現：
+> **arena 大小會改變 decode 期的 I/O 時序，而這個 prompt 的 top-1/top-2
+> 幾乎並列，於是 argmax 換掉、輸出整段岔開。**
+>
+> 換句話說，`determinism.txt` 那份「兩個預算 md5 相同」的結果，
+> 在本機**無法重現**；能重現的只有「同一預算內 6/6 相同」這個部分
+> （本機兩個 build、兩個預算都成立，合計 24/24）。
+> 要引用時請照這個範圍引用，不要說「RAM 預算不影響輸出」。
+
 **而且分歧點已經定位到**：把兩種設定的 MoE dump 逐一比對（43 層 × 17 次 op call
 = 731 個檔案），**43 個 prefill 呼叫全部 bit-identical、688 個 decode 呼叫全部不同**
 （`results/moe_dump_bitexact.md` §B）。分歧已經出現在 MoE op 的**輸入** `cur` 裡，
@@ -513,6 +632,13 @@ results/
   verify_moe_forward.txt         expert 前向 vs 純 NumPy 的逐筆輸出（10/10 PASS）
   moe_dump_bitexact.md           MoE op 的 bit-exact A/B（387/387 與分歧點定位）
   determinism.txt                決定性探針（同一設定重跑 6 次 × 2 個預算）
+  --- 以下為 2026-10-12 獨立複驗新增（原始結果未被覆蓋） ---
+  memory_levels_dsv4-4-8-16-32g-recheck.json / .md
+                                 用 4.2 可攜 ISA 選項重跑整個 4/8/16/32 GiB 矩陣
+  determinism_recheck.txt        決定性探針 × 2 種 build（證明編譯選項非成因）
+  moe_dump_recheck_prompt_dependence.md
+                                 6.3 的 387/387 為何是 prompt 專屬（含對照表）
+  verify_dequant_recheck.txt     修正後 verify_dequant_dsv4.py 的 15/15 輸出
 ```
 
 ### 重跑驗證
@@ -526,10 +652,24 @@ python3 sdq-dsv4/verify_dequant_dsv4.py
 # (2) expert 前向（先跑 sdq-chat --selftest 產生 /tmp/st_l*e*.bin，見 6.2）
 DSV4_LIMIT=10 DSV4_MODE=1 python3 sdq-dsv4/verify_moe_dsv4.py
 
-# (3) MoE op bit-exact A/B（見 6.3 的 SDQ_MOE_DUMP 用法）
+# (3) 決定性探針（同一設定重跑 N 次比對輸出）
 bash sdq-dsv4/determinism_probe.sh 8192 32768
+
+# (4) MoE op bit-exact A/B：同一設定、只改 --ram-budget-mb，各 dump 一次再逐檔 cmp
+#     見 6.3 / results/moe_dump_bitexact.md
+for mb in 4096 32768; do
+  rm -rf /tmp/moe_$mb && mkdir -p /tmp/moe_$mb
+  SDQ_MOE_DUMP=/tmp/moe_$mb/d SDQ_RESERVE_MB=2048 \
+  ./build/bin/sdq-chat -m $MODEL \
+    -p "Explain mixture-of-experts routing in one sentence." -n 8 -c 2048 -t 16 -b 128 \
+    --temp 0 --ram-budget-mb $mb --no-interactive >/dev/null 2>&1
+done
+diff -r --brief /tmp/moe_4096 /tmp/moe_32768 && echo "387/387 bit-identical"
 ```
 
+> ⚠️ **(4) 的結論強烈依賴 prompt**，請見 6.3 節的限定表：換成
+> `The capital of France is` 只會得到 prefill 43/43、decode 0/344。
+
 ---
 
 ## 8. 授權

results/determinism_recheck.txt

@@ -0,0 +1,62 @@
+### determinism — re-verification 2026-10-12
+
+```
+### build A: portable ISA flags (README section 4.2)
+=== ram-budget 8192 MB × 6 ===
+  run 1: 08c061788823
+  run 2: 08c061788823
+  run 3: 08c061788823
+  run 4: 08c061788823
+  run 5: 08c061788823
+  run 6: 08c061788823
+  distinct outputs: 1 / 6
+          6 08c061788823b08e328fd862bf0af9af
+=== ram-budget 32768 MB × 6 ===
+  run 1: 175b890dddeb
+  run 2: 175b890dddeb
+  run 3: 175b890dddeb
+  run 4: 175b890dddeb
+  run 5: 175b890dddeb
+  run 6: 175b890dddeb
+  distinct outputs: 1 / 6
+          6 175b890dddebb91cce0bd2cd9c28e193
+
+### build B: -march=native -mno-amx-int8 -mno-amx-tile -mno-amx-bf16
+=== ram-budget 8192 MB × 6 ===
+  run 1: d5cf2396b424
+  run 2: d5cf2396b424
+  run 3: d5cf2396b424
+  run 4: d5cf2396b424
+  run 5: d5cf2396b424
+  run 6: d5cf2396b424
+  distinct outputs: 1 / 6
+          6 d5cf2396b4240a393092d3a2b2456900
+=== ram-budget 32768 MB × 6 ===
+  run 1: 54a418d53304
+  run 2: 54a418d53304
+  run 3: 54a418d53304
+  run 4: 54a418d53304
+  run 5: 54a418d53304
+  run 6: 54a418d53304
+  distinct outputs: 1 / 6
+          6 54a418d533041aa9af313eff8b01bc33
+```
+
+#### What this shows
+
+1. Within a single fixed configuration the greedy (--temp 0) output is
+   reproducible: 6/6 identical runs at 8192 MB and 6/6 at 32768 MB, in BOTH
+   builds (24/24). This reproduces results/determinism.txt.
+
+2. NEW, NOT reproducible: the published determinism.txt reports the SAME
+   md5 for 8192 MB and 32768 MB (0e637837ac89). Here the two budgets differ
+   in both builds. So --ram-budget-mb belongs on the list of knobs that
+   perturb decode output (README section 6.4), which previously omitted it.
+
+3. The compile flags were the obvious suspect and are EXCLUDED: builds A and B
+   disagree with each other too, and neither reproduces 0e637837ac89.
+
+4. Content at 8192 MB: 'capital of France is **Paris**.'
+   Content at 32768 MB: "'m sorry, I can't answer that question."
+   Consistent with the README 6.4 analysis: this prompt's top-1/top-2 are
+   nearly tied, so the argmax flips. Same config -> same answer, always.

results/memory_levels_dsv4-4-8-16-32g-recheck.json

@@ -0,0 +1,247 @@
+{
+  "model_id": "DeepSeek-V4-Flash-0731-UD-IQ4_NL",
+  "model_repo": "unsloth/DeepSeek-V4-Flash-0731-GGUF",
+  "model_subdir": "UD-IQ4_NL",
+  "architecture": "deepseek4",
+  "shards": 4,
+  "model_bytes": 136662446656,
+  "host": {
+    "cpu_model": "Intel(R) Xeon(R) Platinum 8559C",
+    "cpu_threads": 192,
+    "mem_total_bytes": 2147448107008
+  },
+  "method": "RAM budget passed to the SDQ expert pager arena sizing; no hard cgroup limit (host /sys/fs/cgroup is read-only). Page cache dropped before each run where permitted.",
+  "settings": {
+    "levels_gib": [
+      4,
+      8,
+      16,
+      32
+    ],
+    "repeat": 2,
+    "ctx": 2048,
+    "ubatch": 128,
+    "threads": 16,
+    "n_predict": 48,
+    "reserve_mb": 2048,
+    "temp": 0
+  },
+  "results": [
+    {
+      "level_mb": 4096,
+      "level_gib": 4,
+      "repeat": 1,
+      "prompt": "The capital of France is",
+      "wall_s": 91.45,
+      "decode_tok_s": 0.6,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.6,
+      "prefill_tok": 22,
+      "peak_rss_mb": 2232,
+      "final_rss_mb": 2232,
+      "vmswap_mb": 0,
+      "arena_gib": 1.79,
+      "n_slots": 164,
+      "total_ssd_gb": 158.76,
+      "decode_ssd_mb": 138385.1,
+      "ssd_mb_per_token": 2944.4,
+      "expert_requests": 14551,
+      "expert_hits": 0,
+      "hot_rate_pct": 0,
+      "cold_rate_pct": 100.0,
+      "evictions": 14387,
+      "per_layer_hit_min": 0,
+      "per_layer_hit_max": 0,
+      "assistant": "capital of France is **Paris**.<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition()}. If omitted, the default action is * {@code forward}. For use with custom listener. For more information, in"
+    },
+    {
+      "level_mb": 4096,
+      "level_gib": 4,
+      "repeat": 2,
+      "prompt": "請用一句話說明什麼是 SSD 快取。",
+      "wall_s": 84.45,
+      "decode_tok_s": 0.69,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.7,
+      "prefill_tok": 29,
+      "peak_rss_mb": 2233,
+      "final_rss_mb": 2233,
+      "vmswap_mb": 0,
+      "arena_gib": 1.79,
+      "n_slots": 164,
+      "total_ssd_gb": 165.11,
+      "decode_ssd_mb": 138388.2,
+      "ssd_mb_per_token": 2944.4,
+      "expert_requests": 15133,
+      "expert_hits": 0,
+      "hot_rate_pct": 0,
+      "cold_rate_pct": 100.0,
+      "evictions": 14969,
+      "per_layer_hit_min": 0,
+      "per_layer_hit_max": 0,
+      "assistant": "SSD 快取是使用固態硬碟作為高速暫存區，將頻繁存取的資料預先儲存以加速系統回應的技術。<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition"
+    },
+    {
+      "level_mb": 8192,
+      "level_gib": 8,
+      "repeat": 1,
+      "prompt": "The capital of France is",
+      "wall_s": 52.17,
+      "decode_tok_s": 1.31,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.6,
+      "prefill_tok": 22,
+      "peak_rss_mb": 6325,
+      "final_rss_mb": 6325,
+      "vmswap_mb": 0,
+      "arena_gib": 5.79,
+      "n_slots": 531,
+      "total_ssd_gb": 93.86,
+      "decode_ssd_mb": 71924.1,
+      "ssd_mb_per_token": 1530.3,
+      "expert_requests": 14551,
+      "expert_hits": 5946,
+      "hot_rate_pct": 40.8632,
+      "cold_rate_pct": 59.1368,
+      "evictions": 8074,
+      "per_layer_hit_min": 3.14136,
+      "per_layer_hit_max": 55.0898,
+      "assistant": "capital of France is **Paris**.<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition()}. If omitted, the default action is * {@code forward}. For use with custom listener. For more information, in"
+    },
+    {
+      "level_mb": 8192,
+      "level_gib": 8,
+      "repeat": 2,
+      "prompt": "請用一句話說明什麼是 SSD 快取。",
+      "wall_s": 50.58,
+      "decode_tok_s": 1.5,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.8,
+      "prefill_tok": 29,
+      "peak_rss_mb": 6325,
+      "final_rss_mb": 6325,
+      "vmswap_mb": 0,
+      "arena_gib": 5.79,
+      "n_slots": 531,
+      "total_ssd_gb": 92.34,
+      "decode_ssd_mb": 63877.3,
+      "ssd_mb_per_token": 1359.1,
+      "expert_requests": 15133,
+      "expert_hits": 6663,
+      "hot_rate_pct": 44.0296,
+      "cold_rate_pct": 55.9704,
+      "evictions": 7939,
+      "per_layer_hit_min": 3.64964,
+      "per_layer_hit_max": 62.4277,
+      "assistant": "SSD 快取是使用固態硬碟作為高速暫存區，將頻繁存取的資料預先儲存以加速系統回應的技術。<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition"
+    },
+    {
+      "level_mb": 16384,
+      "level_gib": 16,
+      "repeat": 1,
+      "prompt": "The capital of France is",
+      "wall_s": 53.04,
+      "decode_tok_s": 1.49,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.7,
+      "prefill_tok": 22,
+      "peak_rss_mb": 14507,
+      "final_rss_mb": 14507,
+      "vmswap_mb": 0,
+      "arena_gib": 13.79,
+      "n_slots": 1264,
+      "total_ssd_gb": 71.54,
+      "decode_ssd_mb": 49075.9,
+      "ssd_mb_per_token": 1044.2,
+      "expert_requests": 14551,
+      "expert_hits": 7987,
+      "hot_rate_pct": 54.8897,
+      "cold_rate_pct": 45.1103,
+      "evictions": 5300,
+      "per_layer_hit_min": 15.1832,
+      "per_layer_hit_max": 69.906,
+      "assistant": "'m sorry, I can't answer that question. I'm an AI assistant designed to provide helpful and harmless responses.<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition()}. If omitted, the default act"
+    },
+    {
+      "level_mb": 16384,
+      "level_gib": 16,
+      "repeat": 2,
+      "prompt": "請用一句話說明什麼是 SSD 快取。",
+      "wall_s": 51.11,
+      "decode_tok_s": 1.75,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.8,
+      "prefill_tok": 29,
+      "peak_rss_mb": 14507,
+      "final_rss_mb": 14507,
+      "vmswap_mb": 0,
+      "arena_gib": 13.79,
+      "n_slots": 1264,
+      "total_ssd_gb": 74.52,
+      "decode_ssd_mb": 45626.5,
+      "ssd_mb_per_token": 970.8,
+      "expert_requests": 15133,
+      "expert_hits": 8298,
+      "hot_rate_pct": 54.8338,
+      "cold_rate_pct": 45.1662,
+      "evictions": 5571,
+      "per_layer_hit_min": 16.3415,
+      "per_layer_hit_max": 66.474,
+      "assistant": "SSD 快取是使用固態硬碟作為高速暫存區，將頻繁存取的資料預先儲存以加速系統回應的技術。<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition"
+    },
+    {
+      "level_mb": 32768,
+      "level_gib": 32,
+      "repeat": 1,
+      "prompt": "The capital of France is",
+      "wall_s": 56.59,
+      "decode_tok_s": 1.88,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.8,
+      "prefill_tok": 22,
+      "peak_rss_mb": 30894,
+      "final_rss_mb": 30894,
+      "vmswap_mb": 0,
+      "arena_gib": 29.79,
+      "n_slots": 2732,
+      "total_ssd_gb": 55.6,
+      "decode_ssd_mb": 32746.2,
+      "ssd_mb_per_token": 696.7,
+      "expert_requests": 14551,
+      "expert_hits": 9453,
+      "hot_rate_pct": 64.9646,
+      "cold_rate_pct": 35.035399999999996,
+      "evictions": 2366,
+      "per_layer_hit_min": 41.3613,
+      "per_layer_hit_max": 71.5569,
+      "assistant": "<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition()}. If omitted, the default action is * {@code forward}. For use with custom listener. For more information, including * the valid values of"
+    },
+    {
+      "level_mb": 32768,
+      "level_gib": 32,
+      "repeat": 2,
+      "prompt": "請用一句話說明什麼是 SSD 快取。",
+      "wall_s": 66.17,
+      "decode_tok_s": 1.56,
+      "decode_tok": 47,
+      "prefill_tok_s": 1.9,
+      "prefill_tok": 29,
+      "peak_rss_mb": 30894,
+      "final_rss_mb": 30894,
+      "vmswap_mb": 0,
+      "arena_gib": 29.79,
+      "n_slots": 2732,
+      "total_ssd_gb": 61.14,
+      "decode_ssd_mb": 31923.0,
+      "ssd_mb_per_token": 679.2,
+      "expert_requests": 15133,
+      "expert_hits": 9526,
+      "hot_rate_pct": 62.9485,
+      "cold_rate_pct": 37.0515,
+      "evictions": 2875,
+      "per_layer_hit_min": 28.5366,
+      "per_layer_hit_max": 70.5202,
+      "assistant": "SSD 快取是使用固態硬碟作為高速暫存區，將頻繁存取的資料預先儲存以加速系統回應的技術。<｜end▁of▁sentence｜> of * {@link ListenerRule#getCondition"
+    }
+  ]
+}

results/memory_levels_dsv4-4-8-16-32g-recheck.md

@@ -0,0 +1,24 @@
+<!--
+獨立複驗（2026-10-12），非原始發布結果。
+與 results/memory_levels_dsv4-4-8-16-32g.md 同一支 runner、同一台機器，
+差別只有編譯選項：這一組是用 README 4.2 的可攜 ISA 選項
+（-DGGML_NATIVE=OFF -DGGML_AVX512=ON ...）建的。
+逐欄對照見 README 第 3 節。原始發布結果未被覆蓋，仍在原檔名。
+-->
+
+# RAM 預算矩陣：dsv4-4-8-16-32g
+
+RAM budget passed to the SDQ expert pager arena sizing; no hard cgroup limit (host /sys/fs/cgroup is read-only). Page cache dropped before each run where permitted.
+
+- Host: Intel(R) Xeon(R) Platinum 8559C, 192 threads
+- ctx=2048, ubatch=128, threads=16, output tokens=48, temp=0, repeats=2
+- KV/SSM state + compute reserve: 2048 MB
+
+| RAM budget | arena (GiB) | slots | decode tok/s | prefill tok/s | hot % | cold % | SSD MiB/tok | total SSD read | peak RSS (MB) | swap | per-layer hit min..max |
+|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| 4 GiB | 1.79 | 164 | 0.65 | 1.65 | 0.0 | 100.0 | 2944.4 | 161.9 GB | 2232 | 0 MB | 0.0..0.0 |
+| 8 GiB | 5.79 | 531 | 1.41 | 1.70 | 42.4 | 57.6 | 1444.7 | 93.1 GB | 6325 | 0 MB | 3.4..58.8 |
+| 16 GiB | 13.79 | 1264 | 1.62 | 1.75 | 54.9 | 45.1 | 1007.5 | 73.0 GB | 14507 | 0 MB | 15.8..68.2 |
+| 32 GiB | 29.79 | 2732 | 1.72 | 1.85 | 64.0 | 36.0 | 688.0 | 58.4 GB | 30894 | 0 MB | 34.9..71.0 |
+
+JSON: `memory_levels_dsv4-4-8-16-32g.json`

results/moe_dump_recheck_prompt_dependence.md

@@ -0,0 +1,34 @@
+### MoE op bit-exact A/B — re-verification 2026-10-12
+
+Identical command in both rows; the ONLY difference is `--ram-budget-mb 4096`
+vs `32768`. Build = portable ISA flags (README 4.2), `-n 8`, `ctx=2048`,
+`ubatch=128`, `-t 16`, `--temp 0`. 43 layers x 9 op calls = 387 files per side.
+
+| prompt | prefill bit-identical | decode bit-identical | total |
+|---|---:|---:|---:|
+| `Explain mixture-of-experts routing in one sentence.` (published section A) | **43 / 43** | **344 / 344** | **387 / 387** |
+| `The capital of France is` | **43 / 43** | **0 / 344** | **43 / 387** |
+
+## Reading
+
+* Row 1 **exactly reproduces** the published `results/moe_dump_bitexact.md`
+  section A (387/387). Section A itself is sound and was not wrong.
+* Row 2 is the correction: 387/387 is **prompt-specific**, not a general
+  property of the pager. With a prompt whose top-1/top-2 are nearly tied,
+  only prefill stays bit-exact and every decode call diverges.
+
+* Row 1 first diverging decode call: none (all 344 identical)
+* Row 2 first diverging decode call: layer 0, decode #1 (`d.0.43`)
+  * expert `ids` identical at that point: **False**
+  * `cur` max|delta| / max|cur| = **2.12**
+
+Row 2's divergence is not floating-point noise: at the very FIRST decode step
+`cur` already differs by a relative max of ~2.1 and the routed expert `ids`
+have switched. Prefill was bit-identical, so the chain is: prefill logits
+differ infinitesimally -> `--temp 0` argmax picks a different near-tied token
+-> every later decode input is a different sequence. This matches the
+README 6.4 conclusion that the instability lives in llama.cpp's decode-phase
+dsv4 CSA/HCA path, not in SDQ's paging.
+
+Consequence for citation: always quote 387/387 **together with the prompt**.
+It is not an end-to-end guarantee.

results/verify_dequant_recheck.txt

@@ -0,0 +1,30 @@
+### dequant bit-exactness — re-verification 2026-10-12
+
+Same host/build as the published results. Output of the FIXED
+verify_dequant_dsv4.py: 5 tensors x 3 rows = 15 unique comparisons.
+
+The pre-fix script printed 21 rows (6 redundant, because TARGETS was one flat
+suffix list applied to BOTH layer 0 and layer 26, and ffn_down_exps is MXFP4 in
+both) and self-reported "21/21". After the fix the script and the published
+results/verify_dequant.txt agree at 15/15.
+
+```
+blk.0.ffn_gate_exps.weight         type=21 row=0     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_gate_exps.weight         type=21 row=1     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_gate_exps.weight         type=21 row=137   max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_up_exps.weight           type=21 row=0     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_up_exps.weight           type=21 row=1     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_up_exps.weight           type=21 row=137   max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_down_exps.weight         type=39 row=0     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_down_exps.weight         type=39 row=1     max|NumPy - ggml| = 0  EXACT
+blk.0.ffn_down_exps.weight         type=39 row=137   max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_gate_exps.weight        type=39 row=0     max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_gate_exps.weight        type=39 row=1     max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_gate_exps.weight        type=39 row=137   max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_down_exps.weight        type=39 row=0     max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_down_exps.weight        type=39 row=1     max|NumPy - ggml| = 0  EXACT
+blk.26.ffn_down_exps.weight        type=39 row=137   max|NumPy - ggml| = 0  EXACT
+
+15/15 bit-exact
+ALL BIT-EXACT
+```

sdq-dsv4/determinism_probe.sh

@@ -8,8 +8,24 @@
 #       本來就不決定
 # 兩者必須分開，否則 README 上寫什麼都沒有意義。
 set -u
-BIN=${BIN:-/root/app/work/llama.cpp/build/bin/sdq-chat}
-MODEL=${MODEL:-/root/app/work/models/UD-IQ4_NL/DeepSeek-V4-Flash-0731-UD-IQ4_NL-00001-of-00004.gguf}
+# 預設值不要綁死某台機器的路徑：先用環境變數，否則在 repo 根目錄往上下找。
+#   BIN   = sdq-chat 執行檔（build/bin/sdq-chat）
+#   MODEL = shard 1（metadata-only 的那個 shard，pager 會自己串另外三個）
+# 三者都可以用環境變數覆寫。
+BIN=${BIN:-}
+MODEL=${MODEL:-}
+if [ -z "$BIN" ]; then
+  BIN=$(ls -1 ./llama.cpp/build/bin/sdq-chat ./build/bin/sdq-chat 2>/dev/null | head -1)
+fi
+if [ -z "$MODEL" ]; then
+  MODEL=$(ls -1 ./models/UD-IQ4_NL/*00001-of-00004.gguf 2>/dev/null | head -1)
+fi
+if [ -z "$BIN" ] || [ ! -x "$BIN" ]; then
+  echo "ERROR: 找不到 sdq-chat，請用 BIN=/path/to/sdq-chat 指定" >&2; exit 2
+fi
+if [ -z "$MODEL" ] || [ ! -f "$MODEL" ]; then
+  echo "ERROR: 找不到 GGUF shard 1，請用 MODEL=/path/to/*00001-of-00004.gguf 指定" >&2; exit 2
+fi
 PROMPT=${PROMPT:-"The capital of France is"}
 NTOK=${NTOK:-16}
 N=${N:-6}

sdq-dsv4/verify_dequant_dsv4.py

@@ -24,12 +24,18 @@ import verify_moe_dsv4 as V  # noqa: E402
 GGML_DEQ = os.environ.get('GGML_DEQ', '/tmp/opencode/ggml_deq')
 
 # (tensor suffix, ggml type id, 每列的元素數)
+# 每個項目都是「(層, 張量後綴, 預期 ggml 型別 id, 該列元素數)」。
+#
+# 這裡必須逐層寫死，不能寫成「一張 suffix 清單 + 外層 for il in (0, 26)」：
+# 那樣子 ('ffn_down_exps.weight', 39, ...) 會同時被套用到 layer 0 與 layer 26，
+# 而 layer 0 / layer 26 的 down_exps 都是 MXFP4(39) → 同一組位元組被比對兩次，
+# 於是印出 21 列但只有 15 列是獨立比對（README 6.1 記錄的 6 列重複）。
 TARGETS = [
-    ('ffn_gate_exps.weight', 21, 4096),   # IQ3_S（layer 0）
-    ('ffn_up_exps.weight', 21, 4096),     # IQ3_S（layer 0）
-    ('ffn_down_exps.weight', 39, 2048),   # MXFP4（layer 0）
-    ('ffn_gate_exps.weight', 39, 4096),   # MXFP4（layer 26 的 gate/up）
-    ('ffn_down_exps.weight', 39, 2048),   # MXFP4（layer 26 的 down）
+    (0, 'ffn_gate_exps.weight', 21, 4096),    # IQ3_S（layer 0）
+    (0, 'ffn_up_exps.weight', 21, 4096),      # IQ3_S（layer 0）
+    (0, 'ffn_down_exps.weight', 39, 2048),    # MXFP4（layer 0）
+    (26, 'ffn_gate_exps.weight', 39, 4096),   # MXFP4（layer 26 的 gate）
+    (26, 'ffn_down_exps.weight', 39, 2048),   # MXFP4（layer 26 的 down）
 ]
 
 
@@ -38,34 +44,39 @@ def main() -> int:
 
     rows = []
     all_ok = True
-    for il in (0, 26):
-        for suffix, tt, n in TARGETS:
-            name = 'blk.%d.%s' % (il, suffix)
-            if name not in tensors:
-                continue
-            si, off, dims, ttt = tensors[name]
-            if ttt != tt:
-                continue
-            nrow = V.row_bytes(n, tt)
-            # 取該 (layer, tensor) 的第 0 個 expert 的前幾列
-            for row in (0, 1, 137):
-                byte_off = off + row * nrow
-                raw = V.SHARDS[si]
-                got = subprocess.run([GGML_DEQ, str(tt), raw, str(byte_off), str(n)],
-                                     capture_output=True, check=True).stdout
-                ref = np.frombuffer(got, dtype=np.float32)
-                # ggml_deq 讀的是整列（一個 weight 向量），NumPy 這邊用同一段
-                # 位元組還原 —— 兩邊必須對齊同一個 offset。
-                with open(raw, 'rb') as f:
-                    f.seek(byte_off)
-                    mine = V.dequant(f.read(nrow), n, tt)
-                assert ref.shape == mine.shape == (n,)
-                d = np.abs(ref - mine)
-                exact = bool(np.array_equal(ref, mine))
-                all_ok &= exact
-                rows.append((il, suffix, tt, row, float(d.max()), exact))
-                print('%-34s type=%-2d row=%-4d  max|NumPy - ggml| = %g  %s'
-                      % (name, tt, row, d.max(), 'EXACT' if exact else 'MISMATCH'))
+    for il, suffix, tt, n in TARGETS:
+        name = 'blk.%d.%s' % (il, suffix)
+        if name not in tensors:
+            raise SystemExit('tensor missing from GGUF: %s' % name)
+        si, off, dims, ttt = tensors[name]
+        if ttt != tt:
+            raise SystemExit('%s: expected ggml type %d, header says %d'
+                             % (name, tt, ttt))
+        nrow = V.row_bytes(n, tt)
+        # 取該 (layer, tensor) 的第 0 個 expert 的前幾列
+        for row in (0, 1, 137):
+            byte_off = off + row * nrow
+            raw = V.SHARDS[si]
+            got = subprocess.run([GGML_DEQ, str(tt), raw, str(byte_off), str(n)],
+                                 capture_output=True, check=True).stdout
+            ref = np.frombuffer(got, dtype=np.float32)
+            # ggml_deq 讀的是整列（一個 weight 向量），NumPy 這邊用同一段
+            # 位元組還原 —— 兩邊必須對齊同一個 offset。
+            with open(raw, 'rb') as f:
+                f.seek(byte_off)
+                mine = V.dequant(f.read(nrow), n, tt)
+            assert ref.shape == mine.shape == (n,)
+            d = np.abs(ref - mine)
+            exact = bool(np.array_equal(ref, mine))
+            all_ok &= exact
+            rows.append((il, suffix, tt, row, float(d.max()), exact))
+            print('%-34s type=%-2d row=%-4d  max|NumPy - ggml| = %g  %s'
+                  % (name, tt, row, d.max(), 'EXACT' if exact else 'MISMATCH'))
+
+    seen = {(r[0], r[1], r[2]) for r in rows}
+    if len(seen) != len(TARGETS):
+        raise SystemExit('duplicate comparisons: %d rows over %d unique tensors'
+                         % (len(rows), len(seen)))
 
     print()
     print('%d/%d bit-exact' % (sum(1 for r in rows if r[5]), len(rows)))