diff --git a/README.md b/README.md
index 760fc3bc652de5fbe63c4e461673b4d1014e0234..e944b94465e159385eaf2b5671f5fe0f2ef2ae6b 100644
--- a/README.md
+++ b/README.md
@@ -14,6 +14,13 @@ tags:
 
 CPU expert paging for the Unsloth `flint-UD-Q4_K_XL.gguf` checkpoint. Routed gate/up/down expert weights stay on disk and are copied into a bounded RAM arena as needed. Flint's router, selected-expert weights, shared expert, attention and dense layers remain on the upstream llama.cpp graph.
 
+Two versions are measured throughout:
+
+- **v1**: expert paging with the per-expert custom op ("mode C"). `SDQ_PRELOAD=0 SDQ_ROW_PARALLEL=0`.
+- **v2**: v1 plus preload of all experts into the arena and row-parallel decode for gated (gate/up/down) experts. `SDQ_PRELOAD=1 SDQ_ROW_PARALLEL=1`, arena ≥ 28.6 GiB.
+
+Native llama.cpp (`SDQ_PAGER=0`) is the reference.
+
 ## Source model
 
 - GGUF: [`unsloth/flint-GGUF`](https://huggingface.co/unsloth/flint-GGUF), `flint-UD-Q4_K_XL.gguf` (25,054,688,704 bytes)
@@ -28,13 +35,13 @@ System information recorded on 2026-10-23:
 |---|---|
 | CPU | Intel Xeon Platinum 8559C, 2 sockets × 48 cores, SMT enabled (192 host logical CPUs) |
 | CPU features | AVX2, AVX-512, VNNI, BF16; AMX available on host but disabled for this build |
-| Container CPU quota | 16 CPUs (`cpu.max` = 1,600,000 / 100,000); benchmarks used 4 threads |
+| Container CPU quota | 16 CPUs (`cpu.max` = 1,600,000 / 100,000); v1 memory tiers used 16 threads, v2 and the throughput runs used 16 threads |
 | Host memory | 2.0 TiB reported by `free` |
 | Container memory limit | 104,000,000,000 bytes (about 96.9 GiB), cgroup v2 |
 | Container swap limit | 2,936,848,384 bytes (about 2.74 GiB); measured test processes used 0 swap |
 | GPU | No GPU used; CPU-only build and inference |
 | OS / kernel | Amazon Linux 2023 x86_64, Linux 6.18.48 |
-| llama.cpp | `970b51ce7c9ba5fff0726c5e9994f1c2f0e4eedc` (`970b51c`) + `flint-epq.patch` |
+| llama.cpp | `970b51ce7c9ba5fff0726c5e9994f1c2f0e4eedc` (`970b51c`) + `flint-epq.patch` (sha256 `15c6554271d0c77357d96e628546f8ba624e6834c4c90c69615c90c2ac2310ea`) |
 
 ## Build
 
@@ -53,77 +60,100 @@ cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=OFF \
 cmake --build build -j 8 --target llama-cli llama-server llama-completion llama-perplexity
 ```
 
-Download the GGUF and run the smoke test exactly as it was verified (peak process-tree RSS 2.39 GiB, swap 0, answer produced):
+v1 smoke test (verified; peak process-tree RSS 2.39 GiB, swap 0, answer produced):
 
 ```bash
 hf download unsloth/flint-GGUF flint-UD-Q4_K_XL.gguf --local-dir models
-SDQ_PAGER=1 SDQ_ARENA_MB=512 SDQ_PREFETCH=0 SDQ_PRELOAD=0 \
+SDQ_PAGER=1 SDQ_ARENA_MB=512 SDQ_PREFETCH=0 SDQ_PRELOAD=0 SDQ_ROW_PARALLEL=0 \
   build/bin/llama-cli -m models/flint-UD-Q4_K_XL.gguf -c 2048 \
   -p 'What is 1 + 1? Answer briefly.' -n 32 -t 16 --temp 0 --single-turn --simple-io
 ```
 
-Set `SDQ_PAGER=0` to use the native llama.cpp MoE path. `SDQ_ARENA_MB` sets the expert cache size; `SDQ_PREFETCH=0` is the tested default. The pager emits JSON statistics when `SDQ_STATS_FILE=/path/to/stats.jsonl` is set.
+v2 (preload + row-parallel). The arena must hold all 12,032 experts (≥ 28.6 GiB). Row-parallel applies to decode steps where the number of needed experts is below the thread count, so use at least 7 threads:
+
+```bash
+SDQ_PAGER=1 SDQ_PREFETCH=0 SDQ_ARENA_MB=29400 SDQ_PRELOAD=1 SDQ_ROW_PARALLEL=1 \
+  build/bin/llama-cli -m models/flint-UD-Q4_K_XL.gguf -c 512 -b 128 -ub 32 \
+  -p 'Explain in one sentence why the sky is blue.' -n 32 -t 16 --temp 0 --single-turn --simple-io
+```
+
+`SDQ_PAGER=0` selects the native llama.cpp MoE path. `SDQ_ARENA_MB` sets the expert cache size. `SDQ_STATS_FILE=/path/to/stats.jsonl` writes JSON statistics.
 
 ## Numerical validation
 
-EPQ is compared against native llama.cpp (`SDQ_PAGER=0`) on the same binary. Reference text: the first 6,000 bytes of llama.cpp's `README.md` (`results/logs/ppl_corpus.txt`), 4 chunks of 256 tokens.
+Reference: native llama.cpp (`SDQ_PAGER=0`) on the same binary. Text: the first 6,000 bytes of llama.cpp's `README.md` (`results/logs/common/ppl_corpus.txt`). All logs are in `results/logs/{native,v1,v2}/`.
 
-| Test | Result |
-|---|---|
-| Greedy generation, 24 tokens, `The capital of France is` | Identical to native (`results/logs/greedy_*.txt`) |
-| Perplexity, native | 10.7790 ± 1.6026 |
-| Perplexity, EPQ | 10.8014 ± 1.6076 (+0.21 %) |
-| KL divergence vs native, EPQ | mean 0.01912 ± 0.00198; median 0.00429; max 0.455 |
-| Top-1 token agreement vs native, EPQ | 92.3 % ± 1.2 %; RMS Δp 2.90 % |
-| **Control:** native with ubatch 16 vs native with ubatch 64 (only float summation order changes) | mean KLD 0.02401 ± 0.00339; max 1.136; top-1 agreement 93.1 % |
-| Layer-wise MoE output, layers 2–6, EPQ vs native (same input) | max relative difference ≤ 1.6e-5 |
+| Test | Native | v1 | v2 |
+|---|---|---|---|
+| Greedy, 24 tokens, 8 threads (`The capital of France is`) | reference | identical to native | identical to native |
+| Greedy, 24 tokens, 16 threads | reference | differs from native at char 13 | identical to v1 |
+| Prefill perplexity (`-ub 64`, 8 threads, 4 chunks × 256 tokens) | 10.7790 ± 1.6026 | 10.8014 ± 1.6076 (+0.21 %) | 10.8014 ± 1.6076 |
+| Prefill KL vs native (mean / median / max) | — | 0.019120 / 0.004286 / 0.455 | 0.019120 / 0.004286 / 0.455 |
+| Prefill top-1 agreement vs native | — | 92.3 % | 92.3 % |
+| Decode KL vs native (`-ub 1`, 16 threads, 2 chunks) | — | 0.044057; top-1 92.5 % | 0.044057; top-1 92.5 % |
+| **Control:** native `-ub 64` vs native `-ub 1` (only float order changes) | — | — | mean KLD 0.034464; top-1 90.2 % |
+| Decode v2 vs v1 (`-ub 1`, 16 threads) | — | — | KLD 1e-6 (max 3.8e-5); top-1 100 % |
+| Layer-wise MoE output, layers 2–6, EPQ vs native (same input) | — | max relative diff ≤ 1.6e-5 | — |
 
-The EPQ result is **not bit-identical** to native. The EPQ MoE sums the selected experts in expert-id order, while upstream sums in top-k order, so the outputs differ at the float-rounding level (~1e-7). Activation quantization to 8 bits then amplifies this small difference through the layers, so the difference grows with depth (layers 7 and later). The ubatch control shows that native llama.cpp itself gives a KL divergence and top-1 disagreement of the same size when only float summation order changes. The EPQ deviation is therefore in the same range as ordinary reordering sensitivity for this checkpoint. This is a measured conclusion from these runs, not a proof of exact equivalence.
+Findings:
 
-## 2/4/8/16/32 GiB memory benchmark
+- **v2 row-parallel is numerically equivalent to v1** in decode (KL 1e-6, top-1 100 %) and identical to v1 in prefill (prefill does not use row-parallel). It is not bit-identical.
+- **EPQ is not bit-identical to native.** Float summation order differs, and 8-bit activation quantization amplifies the difference through the layers. The native `-ub 1` vs `-ub 64` control gives a KL divergence of the same magnitude (0.034 vs 0.044). The deviation is in the range of ordinary reordering sensitivity for this checkpoint; it is a measured conclusion from these runs, not a proof of exact equivalence.
+- **Greedy text identity is fragile.** At 8 threads the 24-token output matches native exactly; at 16 threads it diverges after 12 characters, for v1 and v2 alike. The divergence is a near-tie argmax flip caused by the reordering above.
+
+## Memory benchmark (RSS watchdog, v1)
 
 `bench.sh <GiB> <pager 0|1>` runs a fixed prompt and samples process-tree RSS every 20 ms. If sampled RSS exceeds the requested limit, the watchdog kills the run. This is a sampled RSS watchdog, not a kernel/cgroup hard limit: RSS may briefly exceed the threshold between samples, and it counts file-backed pages as reported by `/proc`. `tools/runmem_cgroup.py` (cgroup-v2 `MemoryMax`) is included but **untested in this environment**; it requires a host with writable cgroup v2 and systemd.
 
 ```bash
-for cap in 2 4 8 16 32; do ./bench.sh "$cap" 1; done
+for cap in 2 4 8 16 32; do ./bench.sh "$cap" 1; done   # THREADS defaults to 16
 ```
 
-Override model path, arena size and thread count with `MODEL`, `ARENA_MB` and `THREADS`. The workload is short (context 128, batch 32, ubatch 16, 8 generated tokens) and the benchmark sets `SDQ_PRELOAD=0`, so the preload/row-parallel path is not exercised by these numbers. Keep enough memory for non-expert weights, KV/state and compute buffers; a limit below that floor is expected to fail to load.
+Override with `MODEL`, `ARENA_MB`, `THREADS`. Workload: context 128, batch 32, ubatch 16, 8 generated tokens, prefetch off.
 
 ### Results
 
-Each tier was run 3 times with the same checkpoint, prompt (`Explain in one sentence why the sky is blue.`), 4 CPU threads, pager enabled, prefetch disabled. Table values are medians; all 15 runs exited 0 with 0 swap and never exceeded their watchdog threshold. Raw logs: `results/logs/tier_*g_run*.log`, summary `results/logs/tiers_3runs.txt`.
+Each row is the median of 3 runs. All runs exited 0, swap 0, and stayed under their threshold. Logs: `results/logs/v1/mem_v1_*.log`, `results/logs/v2/v2_32g_run*.log`, summaries in `results/logs/common/mem_summary.txt`.
+
+| Tier | Version | Expert arena (slots) | Peak process-tree RSS | Prompt | Generation |
+|---:|---|---:|---:|---:|---:|
+| 2 GiB | v1 | 256 MiB (105) | 2.014 GiB | 6.9 t/s | 3.9 t/s |
+| 4 GiB | v1 | 1 GiB (420) | 2.765 GiB | 7.1 t/s | 4.4 t/s |
+| 8 GiB | v1 | 3 GiB (1,260) | 4.766 GiB | 9.4 t/s | 4.4 t/s |
+| 16 GiB | v1 | 10 GiB (4,201) | 11.767 GiB | 12.5 t/s | 5.2 t/s |
+| 32 GiB | v1 | 18 GiB (7,561 of 12,032) | 19.766 GiB | 12.0 t/s | 5.2 t/s |
+| 32 GiB | v2 | 28.7 GiB (12,061, all experts) | 30.473 GiB | 33.7 t/s | 13.8 t/s |
 
-| RSS watchdog tier | Expert arena (slots) | Peak process-tree RSS (3 runs) | Swap | Prompt (median) | Generation (median) |
-|---:|---:|---:|---:|---:|---:|
-| 2 GiB | 256 MiB (105) | 1.970 GiB | 0 | 5.5 t/s | 3.2 t/s |
-| 4 GiB | 1 GiB (420) | 2.721 GiB | 0 | 5.8 t/s | 3.1 t/s |
-| 8 GiB | 3 GiB (1,260) | 4.721 GiB | 0 | 5.9 t/s | 3.3 t/s |
-| 16 GiB | 10 GiB (4,201) | 11.721 GiB | 0 | 6.5 t/s | 3.8 t/s |
-| 32 GiB | 18 GiB (7,561 of 12,032 experts) | 19.720 GiB | 0 | 6.7 t/s | 3.9 t/s |
+v2 is only possible in the 32 GiB tier, since preloading all experts needs at least 28.6 GiB of arena. Throughput varies by about ±15 % between runs; treat t/s as approximate.
 
-Throughput varied by about ±15 % between runs of the same tier; treat the t/s columns as approximate. These are measurements on a shared virtualized host, not cgroup-enforced hard limits. Repeat on the target machine for representative throughput.
+## Throughput (v1, v2, native)
 
-## Validation status
+Separate workload from the memory tiers: context 512, batch 128, ubatch 32, 32 generated tokens, 16 threads, 3 runs each (median).
 
-See [`results/test-status.md`](results/test-status.md) for the full system and validation record, including what was not tested.
+| Configuration | Prompt | Generation | Peak RSS | Logs |
+|---|---:|---:|---:|---|
+| Native (`SDQ_PAGER=0`) | 54.4 t/s | 21.5 t/s | 24.31 GiB | `results/logs/native/perf_c512_*` |
+| v1, arena 18 GiB | 13.0 t/s | 5.6 t/s | 19.79 GiB | `results/logs/v1/perf_c512_v1_arena18g_*` |
+| v1, arena 28.7 GiB, no preload | 12.3 t/s | 5.5 t/s | 30.51 GiB | `results/logs/v1/perf_c512_v1_arena28g_*` |
+| **v2**, arena 28.7 GiB, preload + row-parallel | **32.4 t/s** | **16.2 t/s** | 30.50 GiB | `results/logs/v2/perf_c512_v2_arena28g_*` |
 
-## Which EPQ version is ported
+v2 generates about 2.9× faster than v1 and about 75 % of native generation speed. The prompt phase is also faster in v2 (likely because the preload removes SSD reads during prefill, which still uses mode C; not separately measured). The v1 and v2 rows at 28.7 GiB have nearly the same RSS because the prompt already touches most experts in both runs.
 
-The source is [HelloSun's Nemotron EPQ](https://huggingface.co/HelloSun/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF_EPQ) with two patches applied in order: `0001-epq-nemotron-h-moe.patch` (EPQ v1) and `epqv2.patch` (EPQ v2 increment). The Flint port applies both, but only these parts are active for Flint:
+## Version coverage
 
-| Feature | Source | Status for Flint |
+| Feature | Version | Status for Flint |
 |---|---|---|
-| Expert paging, SSD→arena, per-expert custom op ("mode C") | v1 | Active; this is what all measurements use |
-| Preload all experts into the arena | v2 | Code present; runs only when the arena holds all 12,032 experts (about 28.6 GiB at 2.44 MiB/slot). No measured tier reaches this, so it is not exercised here |
-| Row-parallel MoE (16-thread row split) | v2 | Inactive: gated only for no-gate (Nemotron) experts; Flint has gate/up/down |
+| Expert paging, SSD→arena, per-expert custom op ("mode C") | v1 | Active in all v1 and v2 runs (prefill always uses it) |
+| Preload all experts into the arena | v2 | Active when the arena holds all 12,032 experts (≥ 28.6 GiB) |
+| Row-parallel MoE, decode (16-row blocks shared across threads) | v2 | Active for gated Flint experts when the thread count exceeds the number of needed experts (decode, ≥ 7 threads). Ported in this version. |
+| Nemotron no-gate / ReLU² path | v1, v2 | Not used by Flint |
 
-The Flint-specific hook routes gated SiLU experts through the pager. Nemotron's no-gate/ReLU-squared path is not used by Flint.
+The Flint hook routes gated SiLU experts through the pager. The v2 row-parallel gated path computes gate and up for the same row block and applies the same SwiGLU function as the mode C path.
 
-Flint's upstream `build_moe_ffn` also applies a power-of-two amax rescale of the down-projection input and undoes it on the output. This rescale is exact in F32 and Q8_K quantization is scale-invariant, so the paged path omits it. The layer-2–6 MoE comparison against native (which includes the rescale) confirms that the omission changes no results beyond float-summation noise.
+Flint's upstream `build_moe_ffn` also applies a power-of-two amax rescale of the down-projection input and undoes it on the output. This rescale is exact in F32 and Q8_K quantization is scale-invariant, so the paged path omits it. The layer-2–6 MoE comparison against native (which includes the rescale) confirms the omission changes nothing beyond float-summation noise.
 
 ## Attribution
 
-The pager is adapted from HelloSun's Nemotron EPQ v1 and v2 (linked above).
+The pager is adapted from HelloSun's Nemotron EPQ [v1 and v2](https://huggingface.co/HelloSun/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF_EPQ).
 
 The GGUF weights remain hosted by Unsloth. This repository contains source, patch, scripts, logs and test notes only.
diff --git a/flint-epq.patch b/flint-epq.patch
index 67f948f11f41da6a9949e6a3e386a01d184b74ab..c452a9d082bd80f50ad4d9a3259a4fa31a04d7ac 100644
--- a/flint-epq.patch
+++ b/flint-epq.patch
@@ -463,10 +463,10 @@ index 0000000..11c6bc5
 +} // namespace sdq
 diff --git a/src/sdq_moe.cpp b/src/sdq_moe.cpp
 new file mode 100644
-index 0000000..38ad235
+index 0000000..14cc23e
 --- /dev/null
 +++ b/src/sdq_moe.cpp
-@@ -0,0 +1,1150 @@
+@@ -0,0 +1,1182 @@
 +// SDQ 被分頁的 MoE FFN（自訂 ggml op）
 +//
 +// 為什麼要自己算：上游 build_moe_ffn() 用 ggml_mul_mat_id 直接對
@@ -798,6 +798,19 @@ index 0000000..38ad235
 +
 +// ---------------------------------------------------------------- 單一 expert 的數學
 +// 從 arena 的槽位記憶體算出一個 expert 對多個 token 的輸出（加權後直接累加到 out）
++// SwiGLU 啟動，與上游 build_moe_ffn 一致（見 llama-graph.cpp 的 LLM_FFN_SILU 分支）：
++// limit > eps 時 up 兩側限制、silu(gate) 只限制上界；否則 silu(g)*u。
++// expert_forward 與 row-parallel 共用，確保兩條路徑數學完全相同。
++static inline float gated_act(float g, float u, float swiglu_limit) {
++    if (swiglu_limit > 1e-6f) {
++        const float uc = std::max(-swiglu_limit, std::min(swiglu_limit, u));
++        float gc = g / (1.0f + std::exp(-g));
++        gc = std::min(swiglu_limit, gc);
++        return gc * uc;
++    }
++    return (g / (1.0f + std::exp(-g))) * u;
++}
++
 +static void expert_forward(const expert_desc & ed, const uint8_t * base,
 +                           const float * cur, int64_t n_embd,
 +                           const std::vector<std::pair<int32_t, int32_t>> & toks,
@@ -874,19 +887,7 @@ index 0000000..38ad235
 +
 +        PROF_T0_x(_td);   // SiLU + 加權求和
 +        for (int64_t r = 0; r < ed.down_ne0; r++) {
-+            const float g = gate[(size_t) r];
-+            const float u = upv[(size_t) r];
-+            float act;
-+            if (swiglu_limit > 1e-6f) {
-+                const float uc = std::max(-swiglu_limit, std::min(swiglu_limit, u));
-+                float gc = g / (1.0f + std::exp(-g));
-+                // 與上游 build_moe_ffn 一致：silu(gate) 只限制上界（ggml_clamp(-INF, limit)）
-+                gc = std::min(swiglu_limit, gc);
-+                act = gc * uc;
-+            } else {
-+                act = (g / (1.0f + std::exp(-g))) * u;
-+            }
-+            h[(size_t) r] = act;
++            h[(size_t) r] = gated_act(gate[(size_t) r], upv[(size_t) r], swiglu_limit);
 +        }
 +        quantize_act(act_down, h.data(), hq, ed.down_ne0);
 +
@@ -1083,10 +1084,13 @@ index 0000000..38ad235
 +    struct ed_info {
 +        const uint8_t * wu = nullptr;
 +        const uint8_t * wd = nullptr;
++        const uint8_t * wg = nullptr;   // gated 模型才有：gate 權重
 +        size_t up_row = 0;
 +        size_t dn_row = 0;
++        size_t g_row = 0;
 +        vec_dot_fn f_up = nullptr;
 +        vec_dot_fn f_down = nullptr;
++        vec_dot_fn f_gate = nullptr;
 +        const expert_desc * ed = nullptr;
 +        int64_t n_te = 0;
 +    };
@@ -1102,9 +1106,16 @@ index 0000000..38ad235
 +        q.dn_row = ggml_row_size(ed.type[part::DOWN], ed.down_ne0);
 +        q.f_up   = dot_for(ed.type[part::UP]);
 +        q.f_down = dot_for(ed.type[part::DOWN]);
++        if (!R->no_gate) {
++            q.g_row  = ggml_row_size(ed.type[part::GATE], ed.ne0);
++            q.f_gate = dot_for(ed.type[part::GATE]);
++        }
 +        if (sl >= 0) {
 +            q.wu = P.slot_ptr(sl) + ed.slot_off[part::UP];
 +            q.wd = P.slot_ptr(sl) + ed.slot_off[part::DOWN];
++            if (!R->no_gate) {
++                q.wg = P.slot_ptr(sl) + ed.slot_off[part::GATE];
++            }
 +        }
 +    }
 +
@@ -1124,6 +1135,9 @@ index 0000000..38ad235
 +            if (!q.wu || !q.f_up) {
 +                continue;
 +            }
++            if (!R->no_gate && (!q.wg || !q.f_gate)) {
++                continue;
++            }
 +            const auto & toks = R->tok_of_expert[(size_t) R->need[(size_t) j]];
 +            for (size_t ti = 0; ti < toks.size(); ti++) {
 +                const int32_t t = toks[ti].first;
@@ -1138,8 +1152,17 @@ index 0000000..38ad235
 +                    float acc = 0.0f;
 +                    dot_row(q.f_up, q.wu + (size_t) r * q.up_row, xq,
 +                            q.up_row, (int) q.ed->ne0, acc);
-+                    acc = acc > 0.0f ? acc : 0.0f;
-+                    h[(size_t) r] = acc * acc;
++                    if (R->no_gate) {
++                        // Nemotron LatentMoE：relu(x)^2，沒有 gate
++                        acc = acc > 0.0f ? acc : 0.0f;
++                        h[(size_t) r] = acc * acc;
++                    } else {
++                        // gated（Flint）：silu(gate)*up，與 expert_forward 相同的 SwiGLU
++                        float g = 0.0f;
++                        dot_row(q.f_gate, q.wg + (size_t) r * q.g_row, xq,
++                                q.g_row, (int) q.ed->ne0, g);
++                        h[(size_t) r] = gated_act(g, acc, R->swiglu_limit);
++                    }
 +                }
 +            }
 +        }
@@ -1309,7 +1332,14 @@ index 0000000..38ad235
 +    if (ith == 0) {
 +        R->use_rowpar = false;
 +        R->rowpar_miss = 0;
-+        if (rowpar_env && R->no_gate && n_need > 0 && n_need < nth) {
++        // gated（Flint）需要 gate 與 up 共用同一個 activation 量化型別，才能共用 xq。
++        bool types_ok = true;
++        if (!R->no_gate && n_need > 0) {
++            const expert_desc & e0 = L->experts[(size_t) R->need[0]];
++            types_ok = !e0.fused_gate_up && dot_for(e0.type[part::GATE]) && dot_for(e0.type[part::UP]) &&
++                       act_type_for(e0.type[part::GATE]) == act_type_for(e0.type[part::UP]);
++        }
++        if (rowpar_env && n_need > 0 && n_need < nth && types_ok) {
 +            int miss = 0;
 +            for (int i = 0; i < n_need; i++) {
 +                if (!P.resident(R->il, R->need[(size_t) i])) {
@@ -1557,8 +1587,10 @@ index 0000000..38ad235
 +ggml_tensor * sdq_build_moe(ggml_cgraph * gf, ggml_context * ctx, ggml_tensor * cur,
 +                            ggml_tensor * ids, ggml_tensor * weights, int il, int64_t n_expert,
 +                            int64_t n_expert_used, float swiglu_limit) {
-+    (void) swiglu_limit;
-+    return sdq_build_moe_ex(gf, ctx, cur, ids, weights, il, n_expert, n_expert_used, false, false);
++    ggml_tensor * out = sdq_build_moe_ex(gf, ctx, cur, ids, weights, il, n_expert, n_expert_used, false, false);
++    // 把 swiglu limit 傳給 op（先前這裡 (void) 丟掉，造成 clamp 永遠不生效）
++    rec_for(il)->swiglu_limit = swiglu_limit;
++    return out;
 +}
 +
 +} // namespace sdq
diff --git a/results/logs/common/mem_summary.txt b/results/logs/common/mem_summary.txt
new file mode 100644
index 0000000000000000000000000000000000000000..caa22a34f76dcc517cc48f01c64767033150cce7
--- /dev/null
+++ b/results/logs/common/mem_summary.txt
@@ -0,0 +1,18 @@
+v1 2 GiB run1 Prompt: 7.4 t/s | Generation: 3.9 "peak_gib": 2.014 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 2 GiB run2 Prompt: 6.9 t/s | Generation: 3.8 "peak_gib": 2.013 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 2 GiB run3 Prompt: 6.8 t/s | Generation: 3.9 "peak_gib": 2.014 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 4 GiB run1 Prompt: 7.1 t/s | Generation: 4.4 "peak_gib": 2.765 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 4 GiB run2 Prompt: 7.1 t/s | Generation: 4.7 "peak_gib": 2.765 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 4 GiB run3 Prompt: 7.1 t/s | Generation: 4.4 "peak_gib": 2.765 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 8 GiB run1 Prompt: 8.4 t/s | Generation: 4.5 "peak_gib": 4.765 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 8 GiB run2 Prompt: 9.4 t/s | Generation: 4.4 "peak_gib": 4.766 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 8 GiB run3 Prompt: 9.6 t/s | Generation: 4.4 "peak_gib": 4.766 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 16 GiB run1 Prompt: 12.0 t/s | Generation: 5.2 "peak_gib": 11.767 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 16 GiB run2 Prompt: 13.3 t/s | Generation: 5.2 "peak_gib": 11.766 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 16 GiB run3 Prompt: 12.5 t/s | Generation: 5.4 "peak_gib": 11.767 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 32 GiB run1 Prompt: 12.7 t/s | Generation: 5.2 "peak_gib": 19.766 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 32 GiB run2 Prompt: 11.6 t/s | Generation: 5.1 "peak_gib": 19.766 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v1 32 GiB run3 Prompt: 12.0 t/s | Generation: 5.3 "peak_gib": 19.765 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v2 32 GiB run1 Prompt: 33.7 t/s | Generation: 17.3 "peak_gib": 30.473 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v2 32 GiB run2 Prompt: 30.8 t/s | Generation: 13.7 "peak_gib": 30.473 "swap_kb": 0 "limit_exceeded": false "exit": 0
+v2 32 GiB run3 Prompt: 34.5 t/s | Generation: 13.8 "peak_gib": 30.473 "swap_kb": 0 "limit_exceeded": false "exit": 0
diff --git a/results/logs/common/perf_c512_summary.txt b/results/logs/common/perf_c512_summary.txt
new file mode 100644
index 0000000000000000000000000000000000000000..ff3edc642e364fa4e2d20598266f8aeb8da03fe4
--- /dev/null
+++ b/results/logs/common/perf_c512_summary.txt
@@ -0,0 +1,12 @@
+v1_arena18g run1 Prompt: 12.0 t/s | Generation: 5.5 "peak_gib": 19.794 "swap_kb": 0 "exit": 0
+v1_arena18g run2 Prompt: 13.0 t/s | Generation: 5.6 "peak_gib": 19.794 "swap_kb": 0 "exit": 0
+v1_arena18g run3 Prompt: 13.0 t/s | Generation: 5.7 "peak_gib": 19.794 "swap_kb": 0 "exit": 0
+v1_arena28g run1 Prompt: 11.5 t/s | Generation: 5.5 "peak_gib": 30.507 "swap_kb": 0 "exit": 0
+v1_arena28g run2 Prompt: 12.6 t/s | Generation: 5.5 "peak_gib": 30.507 "swap_kb": 0 "exit": 0
+v1_arena28g run3 Prompt: 12.3 t/s | Generation: 6.2 "peak_gib": 30.507 "swap_kb": 0 "exit": 0
+v2_arena28g run1 Prompt: 31.1 t/s | Generation: 14.5 "peak_gib": 30.499 "swap_kb": 0 "exit": 0
+v2_arena28g run2 Prompt: 32.4 t/s | Generation: 16.2 "peak_gib": 30.499 "swap_kb": 0 "exit": 0
+v2_arena28g run3 Prompt: 40.0 t/s | Generation: 17.3 "peak_gib": 30.498 "swap_kb": 0 "exit": 0
+native run1 Prompt: 61.4 t/s | Generation: 23.3 "peak_gib": 24.315 "swap_kb": 0 "exit": 0
+native run2 Prompt: 54.4 t/s | Generation: 21.5 "peak_gib": 24.311 "swap_kb": 0 "exit": 0
+native run3 Prompt: 44.7 t/s | Generation: 19.8 "peak_gib": 24.31 "swap_kb": 0 "exit": 0
diff --git a/results/logs/ppl_corpus.txt b/results/logs/common/ppl_corpus.txt
similarity index 100%
rename from results/logs/ppl_corpus.txt
rename to results/logs/common/ppl_corpus.txt
diff --git a/results/logs/greedy_epq_fresh.txt b/results/logs/native/greedy_t16.txt
similarity index 100%
rename from results/logs/greedy_epq_fresh.txt
rename to results/logs/native/greedy_t16.txt
diff --git a/results/logs/greedy_native.txt b/results/logs/native/greedy_t8.txt
similarity index 100%
rename from results/logs/greedy_native.txt
rename to results/logs/native/greedy_t8.txt
diff --git a/results/logs/kl_native_ub16_vs_native.log b/results/logs/native/kl_ub16_vs_native_ub64.log
similarity index 100%
rename from results/logs/kl_native_ub16_vs_native.log
rename to results/logs/native/kl_ub16_vs_native_ub64.log
diff --git a/results/logs/native/kl_ub64_vs_native_ub1_t16.log b/results/logs/native/kl_ub64_vs_native_ub1_t16.log
new file mode 100644
index 0000000000000000000000000000000000000000..6c84048913b7347f504705d5f710ede3095f56d9
--- /dev/null
+++ b/results/logs/native/kl_ub64_vs_native_ub1_t16.log
@@ -0,0 +1,50 @@
+0.01.488.946 I cmn          init: llama threadpool init, n_threads = 16
+0.01.544.647 I 
+0.01.544.771 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.562.447 I kl_divergence: computing over 2 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.04.557.676 I kl_divergence: 3.00 seconds per pass - ETA 0.08 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1      13.8921 ±    4.7298       0.03108 ±    0.03658       0.03632 ±    0.01491     3.273 ±  0.529 %    90.551 ±  2.606 %
+   2      14.8507 ±    3.2268       0.01606 ±    0.03056       0.03446 ±    0.00844     2.798 ±  0.338 %    90.157 ±  1.873 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  14.850718 ±   3.226807
+Mean PPL(base)                :  14.614086 ±   3.091134
+Cor(ln(PPL(Q)), ln(PPL(base))):  99.02%
+Mean ln(PPL(Q)/PPL(base))     :   0.016062 ±   0.030563
+Mean PPL(Q)/PPL(base)         :   1.016192 ±   0.031058
+Mean PPL(Q)-PPL(base)         :   0.236632 ±   0.462519
+
+====== KL divergence statistics ======
+Mean    KLD:   0.034464 ±   0.008439
+Maximum KLD:   1.809971
+99.9%   KLD:   1.497865
+99.0%   KLD:   0.475971
+95.0%   KLD:   0.113679
+90.0%   KLD:   0.052229
+Median  KLD:   0.005795
+10.0%   KLD:   0.000036
+ 5.0%   KLD:   0.000006
+ 1.0%   KLD:  -0.000000
+ 0.1%   KLD:  -0.000001
+Minimum KLD:  -0.000001
+
+====== Token probability statistics ======
+Mean    Δp:  0.014 ± 0.176 %
+Maximum Δp: 17.039%
+99.9%   Δp: 15.461%
+99.0%   Δp: 10.002%
+95.0%   Δp:  2.763%
+90.0%   Δp:  1.551%
+75.0%   Δp:  0.226%
+Median  Δp:  0.000%
+25.0%   Δp: -0.155%
+10.0%   Δp: -1.820%
+ 5.0%   Δp: -3.285%
+ 1.0%   Δp: -10.016%
+ 0.1%   Δp: -14.516%
+Minimum Δp: -15.501%
+RMS Δp    :  2.798 ± 0.338 %
+Same top p: 90.157 ± 1.873 %
+
diff --git a/results/logs/native/perf_c512_native_run1.log b/results/logs/native/perf_c512_native_run1.log
new file mode 100644
index 0000000000000000000000000000000000000000..7138ddbac9e59fa86eaac03d4bda2581caf8c6bd
--- /dev/null
+++ b/results/logs/native/perf_c512_native_run1.log
@@ -0,0 +1,41 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+
+[ Prompt: 61.4 t/s | Generation: 23.3 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 25495692, "peak_gib": 24.315, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 63.97, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/native/perf_c512_native_run2.log b/results/logs/native/perf_c512_native_run2.log
new file mode 100644
index 0000000000000000000000000000000000000000..f413f9807cf0b9688f26678e71dded2cc42235db
--- /dev/null
+++ b/results/logs/native/perf_c512_native_run2.log
@@ -0,0 +1,41 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+
+[ Prompt: 54.4 t/s | Generation: 21.5 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 25491508, "peak_gib": 24.311, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 3.74, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/native/perf_c512_native_run3.log b/results/logs/native/perf_c512_native_run3.log
new file mode 100644
index 0000000000000000000000000000000000000000..13b54bc556f5bdb68b29dbc398bea460760140cb
--- /dev/null
+++ b/results/logs/native/perf_c512_native_run3.log
@@ -0,0 +1,41 @@
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+
+[ Prompt: 44.7 t/s | Generation: 19.8 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 25490932, "peak_gib": 24.31, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 3.96, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/native/ppl_kl_base_decode_ub1_t16.log b/results/logs/native/ppl_kl_base_decode_ub1_t16.log
new file mode 100644
index 0000000000000000000000000000000000000000..4647340da2baabf0dd5084c2aac502991ad86981
--- /dev/null
+++ b/results/logs/native/ppl_kl_base_decode_ub1_t16.log
@@ -0,0 +1,11 @@
+0.01.530.395 I cmn          init: llama threadpool init, n_threads = 16
+0.01.656.534 I 
+0.01.656.668 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.01.656.779 I perplexity: saving all logits to /tmp/opencode/flint-work/kt1/base_native_ub1.bin
+0.01.656.787 I perplexity: tokenizing the input ..
+0.01.666.455 I perplexity: tokenization took 9.663 ms
+0.01.666.471 I perplexity: calculating perplexity over 2 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.11.685.672 I perplexity: 9.99 seconds per pass - ETA 0.32 minutes
+[1]13.6918,[2]14.7355,
+0.21.643.935 I Final estimate: PPL = 14.7355 +/- 3.14946
+
diff --git a/results/logs/ppl_kl_native_base_ub64.log b/results/logs/native/ppl_kl_base_prefill_ub64_t8.log
similarity index 100%
rename from results/logs/ppl_kl_native_base_ub64.log
rename to results/logs/native/ppl_kl_base_prefill_ub64_t8.log
diff --git a/results/logs/tiers_3runs.txt b/results/logs/tiers_3runs.txt
deleted file mode 100644
index 07cd7514ee89078721bd5dc7d22b50aa7487663a..0000000000000000000000000000000000000000
--- a/results/logs/tiers_3runs.txt
+++ /dev/null
@@ -1,15 +0,0 @@
-2 1 Prompt: 6.3 t/s | Generation: 3.2 "peak_gib": 1.97 "swap_kb": 0 "limit_exceeded": false "exit": 0
-2 2 Prompt: 5.5 t/s | Generation: 3.2 "peak_gib": 1.97 "swap_kb": 0 "limit_exceeded": false "exit": 0
-2 3 Prompt: 5.5 t/s | Generation: 3.0 "peak_gib": 1.97 "swap_kb": 0 "limit_exceeded": false "exit": 0
-4 1 Prompt: 6.0 t/s | Generation: 3.1 "peak_gib": 2.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-4 2 Prompt: 5.8 t/s | Generation: 2.8 "peak_gib": 2.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-4 3 Prompt: 5.8 t/s | Generation: 3.3 "peak_gib": 2.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-8 1 Prompt: 5.7 t/s | Generation: 3.3 "peak_gib": 4.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-8 2 Prompt: 5.9 t/s | Generation: 3.5 "peak_gib": 4.721 "swap_kb": 0 "limit_exceeded": false "exit": 0
-8 3 Prompt: 5.9 t/s | Generation: 3.3 "peak_gib": 4.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-16 1 Prompt: 6.3 t/s | Generation: 3.7 "peak_gib": 11.721 "swap_kb": 0 "limit_exceeded": false "exit": 0
-16 2 Prompt: 6.6 t/s | Generation: 3.8 "peak_gib": 11.721 "swap_kb": 0 "limit_exceeded": false "exit": 0
-16 3 Prompt: 6.5 t/s | Generation: 3.9 "peak_gib": 11.721 "swap_kb": 0 "limit_exceeded": false "exit": 0
-32 1 Prompt: 6.7 t/s | Generation: 3.9 "peak_gib": 19.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-32 2 Prompt: 6.7 t/s | Generation: 3.7 "peak_gib": 19.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
-32 3 Prompt: 6.0 t/s | Generation: 3.9 "peak_gib": 19.72 "swap_kb": 0 "limit_exceeded": false "exit": 0
diff --git a/results/logs/v1/greedy_t16.txt b/results/logs/v1/greedy_t16.txt
new file mode 100644
index 0000000000000000000000000000000000000000..84c4fd2f82b06d02f7c64b81cd118201ac404e24
--- /dev/null
+++ b/results/logs/v1/greedy_t16.txt
@@ -0,0 +1,2 @@
+ Paris is a true statement. The capital of France is London is a false. The capital of France is Berlin is a
+
diff --git a/results/logs/v1/greedy_t8.txt b/results/logs/v1/greedy_t8.txt
new file mode 100644
index 0000000000000000000000000000000000000000..f5e3f96f9a1ce6be6680de660e7cff069d078c32
--- /dev/null
+++ b/results/logs/v1/greedy_t8.txt
@@ -0,0 +1,2 @@
+ Paris is a simple true/false question. The answer is true. The capital of France is indeed Paris. The user is
+
diff --git a/results/logs/v1/kl_decode_ub1_t16_vs_native.log b/results/logs/v1/kl_decode_ub1_t16_vs_native.log
new file mode 100644
index 0000000000000000000000000000000000000000..a4dfd26e0523ffa0786e72f4e0dc99200130e67e
--- /dev/null
+++ b/results/logs/v1/kl_decode_ub1_t16_vs_native.log
@@ -0,0 +1,70 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.76 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+0.27.048.482 I cmn          init: llama threadpool init, n_threads = 16
+0.27.205.762 I 
+0.27.205.855 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.27.221.544 I kl_divergence: computing over 2 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.48.384.347 I kl_divergence: 21.16 seconds per pass - ETA 0.70 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1      13.7740 ±    4.6245       0.02254 ±    0.03440       0.03177 ±    0.01493     4.997 ±  2.174 %    95.276 ±  1.890 %
+   2      14.8027 ±    3.1804       0.01283 ±    0.02276       0.04406 ±    0.01523     4.051 ±  1.350 %    92.520 ±  1.654 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  14.802734 ±   3.180425
+Mean PPL(base)                :  14.614086 ±   3.091134
+Cor(ln(PPL(Q)), ln(PPL(base))):  99.44%
+Mean ln(PPL(Q)/PPL(base))     :   0.012826 ±   0.022762
+Mean PPL(Q)/PPL(base)         :   1.012909 ±   0.023056
+Mean PPL(Q)-PPL(base)         :   0.188648 ±   0.342994
+
+====== KL divergence statistics ======
+Mean    KLD:   0.044057 ±   0.015234
+Maximum KLD:   3.161583
+99.9%   KLD:   2.810790
+99.0%   KLD:   0.854633
+95.0%   KLD:   0.086102
+90.0%   KLD:   0.048945
+Median  KLD:   0.005596
+10.0%   KLD:   0.000030
+ 5.0%   KLD:   0.000006
+ 1.0%   KLD:   0.000000
+ 0.1%   KLD:  -0.000003
+Minimum KLD:  -0.000004
+
+====== Token probability statistics ======
+Mean    Δp: -0.309 ± 0.254 %
+Maximum Δp: 12.097%
+99.9%   Δp: 11.854%
+99.0%   Δp:  7.841%
+95.0%   Δp:  2.800%
+90.0%   Δp:  1.211%
+75.0%   Δp:  0.182%
+Median  Δp:  0.000%
+25.0%   Δp: -0.223%
+10.0%   Δp: -2.155%
+ 5.0%   Δp: -3.443%
+ 1.0%   Δp: -9.554%
+ 0.1%   Δp: -43.359%
+Minimum Δp: -52.532%
+RMS Δp    :  4.051 ± 1.350 %
+Same top p: 92.520 ± 1.654 %
+
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.83 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
diff --git a/results/logs/kl_epq_vs_native.log b/results/logs/v1/kl_prefill_ub64_t8_vs_native.log
similarity index 100%
rename from results/logs/kl_epq_vs_native.log
rename to results/logs/v1/kl_prefill_ub64_t8_vs_native.log
diff --git a/results/logs/tier_16g_run3.log b/results/logs/v1/mem_v1_16g_run1.log
similarity index 86%
rename from results/logs/tier_16g_run3.log
rename to results/logs/v1/mem_v1_16g_run1.log
index 52eb2226807431c88df9c1f4e381ae199b2b3401..2f68f0e492655d15eb2b43ef18fe8dd944e905e9 100644
--- a/results/logs/tier_16g_run3.log
+++ b/results/logs/v1/mem_v1_16g_run1.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 4201
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.5 t/s | Generation: 3.9 t/s ]
+[ Prompt: 12.0 t/s | Generation: 5.2 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 12290252, "peak_gib": 11.721, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 12.0, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 12338316, "peak_gib": 11.767, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 10.13, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_16g_run1.log b/results/logs/v1/mem_v1_16g_run2.log
similarity index 83%
rename from results/logs/tier_16g_run1.log
rename to results/logs/v1/mem_v1_16g_run2.log
index 5a4a8ce381ed6ad3ee462a9036b9a1e5633395d2..efe099d24ba4a7ce570eb53c64bd1f81304a4886 100644
--- a/results/logs/tier_16g_run1.log
+++ b/results/logs/v1/mem_v1_16g_run2.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 4201
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.3 t/s | Generation: 3.7 t/s ]
+[ Prompt: 13.3 t/s | Generation: 5.2 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 12290444, "peak_gib": 11.721, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 12.29, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 12338064, "peak_gib": 11.766, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 10.35, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_16g_run2.log b/results/logs/v1/mem_v1_16g_run3.log
similarity index 83%
rename from results/logs/tier_16g_run2.log
rename to results/logs/v1/mem_v1_16g_run3.log
index d893e538f6a9e2c9c79e86583b74505f6485b264..32faf5e40342ba71563f14e33c1a8fd7b53b7e2e 100644
--- a/results/logs/tier_16g_run2.log
+++ b/results/logs/v1/mem_v1_16g_run3.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=4201  arena=10.00 GiB  (RAM 實測 RSS=10.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 4201
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.6 t/s | Generation: 3.8 t/s ]
+[ Prompt: 12.5 t/s | Generation: 5.4 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 12290756, "peak_gib": 11.721, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 12.55, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 12338768, "peak_gib": 11.767, "limit_gib": 16.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 10.08, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=10240", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_2g_run2.log b/results/logs/v1/mem_v1_2g_run1.log
similarity index 86%
rename from results/logs/tier_2g_run2.log
rename to results/logs/v1/mem_v1_2g_run1.log
index a15151023afd117053713e06ad65ba40245899ef..6b65b302eed869f07a00b6d57a6d5cf0e7151533 100644
--- a/results/logs/tier_2g_run2.log
+++ b/results/logs/v1/mem_v1_2g_run1.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.38 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.42 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 105
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.5 t/s | Generation: 3.2 t/s ]
+[ Prompt: 7.4 t/s | Generation: 3.9 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2066160, "peak_gib": 1.97, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.11, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2112124, "peak_gib": 2.014, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 4.98, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_2g_run3.log b/results/logs/v1/mem_v1_2g_run2.log
similarity index 86%
rename from results/logs/tier_2g_run3.log
rename to results/logs/v1/mem_v1_2g_run2.log
index 6e69ec60b5654c6befe069b0e6b08fa8634f27c3..5a8a7482a2a0675cbe2e02f513b9e80b5d103005 100644
--- a/results/logs/tier_2g_run3.log
+++ b/results/logs/v1/mem_v1_2g_run2.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.38 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.42 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 105
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.5 t/s | Generation: 3.0 t/s ]
+[ Prompt: 6.9 t/s | Generation: 3.8 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2065892, "peak_gib": 1.97, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.21, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2110928, "peak_gib": 2.013, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.14, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_2g_run1.log b/results/logs/v1/mem_v1_2g_run3.log
similarity index 86%
rename from results/logs/tier_2g_run1.log
rename to results/logs/v1/mem_v1_2g_run3.log
index bb53056065ca9f5d0d82bcd433c75133aba23468..87650194b51d619c857831568558cf7a2a32503f 100644
--- a/results/logs/tier_2g_run1.log
+++ b/results/logs/v1/mem_v1_2g_run3.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.38 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=105  arena=0.25 GiB  (RAM 實測 RSS=0.42 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 105
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.3 t/s | Generation: 3.2 t/s ]
+[ Prompt: 6.8 t/s | Generation: 3.9 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2065880, "peak_gib": 1.97, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.84, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2111384, "peak_gib": 2.014, "limit_gib": 2.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.11, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=256", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_32g_run2.log b/results/logs/v1/mem_v1_32g_run1.log
similarity index 86%
rename from results/logs/tier_32g_run2.log
rename to results/logs/v1/mem_v1_32g_run1.log
index c8827a66614cdf087d5ce5e800af961cc43fd568..9d27dd8eb68af19ee5a2ff5ceef05cd9a312289f 100644
--- a/results/logs/tier_32g_run2.log
+++ b/results/logs/v1/mem_v1_32g_run1.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 7561
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.7 t/s | Generation: 3.7 t/s ]
+[ Prompt: 12.7 t/s | Generation: 5.2 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 20677612, "peak_gib": 19.72, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 17.28, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 20725676, "peak_gib": 19.766, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 15.36, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_32g_run3.log b/results/logs/v1/mem_v1_32g_run2.log
similarity index 86%
rename from results/logs/tier_32g_run3.log
rename to results/logs/v1/mem_v1_32g_run2.log
index 93bfd88e611a70e90c4bfe857b2217f76071d26b..8662bf6ee2823aa2d9826d9aeaeed6a2593955ec 100644
--- a/results/logs/tier_32g_run3.log
+++ b/results/logs/v1/mem_v1_32g_run2.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 7561
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.0 t/s | Generation: 3.9 t/s ]
+[ Prompt: 11.6 t/s | Generation: 5.1 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 20678068, "peak_gib": 19.72, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 17.72, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 20725876, "peak_gib": 19.766, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 15.45, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_32g_run1.log b/results/logs/v1/mem_v1_32g_run3.log
similarity index 86%
rename from results/logs/tier_32g_run1.log
rename to results/logs/v1/mem_v1_32g_run3.log
index 170c5a681aa5e9e8b33647b4a140e25c54ed65a1..9050cadeefba89193fb7f5c783e936753d75c9ae 100644
--- a/results/logs/tier_32g_run1.log
+++ b/results/logs/v1/mem_v1_32g_run3.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 7561
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.7 t/s | Generation: 3.9 t/s ]
+[ Prompt: 12.0 t/s | Generation: 5.3 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 20677644, "peak_gib": 19.72, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 17.06, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 20725484, "peak_gib": 19.765, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 15.68, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_4g_run2.log b/results/logs/v1/mem_v1_4g_run1.log
similarity index 86%
rename from results/logs/tier_4g_run2.log
rename to results/logs/v1/mem_v1_4g_run1.log
index 7458920c1ecde3a194e0ac4269d1e6b6f7dfbde6..17a8ee1e136d021d592286e31f69f4cdc9b4e2ed 100644
--- a/results/logs/tier_4g_run2.log
+++ b/results/logs/v1/mem_v1_4g_run1.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 420
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.8 t/s | Generation: 2.8 t/s ]
+[ Prompt: 7.1 t/s | Generation: 4.4 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2852268, "peak_gib": 2.72, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.93, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2899620, "peak_gib": 2.765, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.31, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_4g_run3.log b/results/logs/v1/mem_v1_4g_run2.log
similarity index 80%
rename from results/logs/tier_4g_run3.log
rename to results/logs/v1/mem_v1_4g_run2.log
index 4bc2c0f790268cc934ae9d76a75bdff4000e9b5a..de57496454f1be1894ae86551ea54890ec705b90 100644
--- a/results/logs/tier_4g_run3.log
+++ b/results/logs/v1/mem_v1_4g_run2.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 420
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.8 t/s | Generation: 3.3 t/s ]
+[ Prompt: 7.1 t/s | Generation: 4.7 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2852068, "peak_gib": 2.72, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.6, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2899720, "peak_gib": 2.765, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.18, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_4g_run1.log b/results/logs/v1/mem_v1_4g_run3.log
similarity index 80%
rename from results/logs/tier_4g_run1.log
rename to results/logs/v1/mem_v1_4g_run3.log
index a3ea65fd612c7a06dd51ca90278022a6d798610b..60289df520972d8ef44f194b2bcdffa2de830a94 100644
--- a/results/logs/tier_4g_run1.log
+++ b/results/logs/v1/mem_v1_4g_run3.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=420  arena=1.00 GiB  (RAM 實測 RSS=1.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 420
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 6.0 t/s | Generation: 3.1 t/s ]
+[ Prompt: 7.1 t/s | Generation: 4.4 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 2852152, "peak_gib": 2.72, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.6, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 2899752, "peak_gib": 2.765, "limit_gib": 4.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 5.33, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=1024", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_8g_run1.log b/results/logs/v1/mem_v1_8g_run1.log
similarity index 86%
rename from results/logs/tier_8g_run1.log
rename to results/logs/v1/mem_v1_8g_run1.log
index 886840dc4705789721c1458af61e7d5b71cbdd2f..af0fca3b102ca0c4529953f81fc381076de832b8 100644
--- a/results/logs/tier_8g_run1.log
+++ b/results/logs/v1/mem_v1_8g_run1.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 1260
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.7 t/s | Generation: 3.3 t/s ]
+[ Prompt: 8.4 t/s | Generation: 4.5 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 4949192, "peak_gib": 4.72, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 7.79, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 4996628, "peak_gib": 4.765, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.15, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_8g_run3.log b/results/logs/v1/mem_v1_8g_run2.log
similarity index 86%
rename from results/logs/tier_8g_run3.log
rename to results/logs/v1/mem_v1_8g_run2.log
index abd4416e2c04e15843c5f75459ce7b18c9641f44..af6bdb3c5dc6944e60984b60d6b44df09e394148 100644
--- a/results/logs/tier_8g_run3.log
+++ b/results/logs/v1/mem_v1_8g_run2.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 1260
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.9 t/s | Generation: 3.3 t/s ]
+[ Prompt: 9.4 t/s | Generation: 4.4 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 4949056, "peak_gib": 4.72, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 7.86, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 4997400, "peak_gib": 4.766, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.26, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/tier_8g_run2.log b/results/logs/v1/mem_v1_8g_run3.log
similarity index 83%
rename from results/logs/tier_8g_run2.log
rename to results/logs/v1/mem_v1_8g_run3.log
index e72dd5f0a997abee135bd5133af7e55042e1cb98..fadb347141a07e7f3a6cb1e520c24265bf745dc3 100644
--- a/results/logs/tier_8g_run2.log
+++ b/results/logs/v1/mem_v1_8g_run3.log
@@ -43,13 +43,13 @@ The user is asking for a one
 [sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
 [sdq] expert pager ready: /tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf
 [sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
-[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.13 GiB)
+[sdq]   slot=2.44 MiB (2555904 B)  slots=1260  arena=3.00 GiB  (RAM 實測 RSS=3.17 GiB)
 [sdq]   evict_policy=lru  stale_window=640
 [sdq]   槽位：單一尺寸 2.44 MiB × 1260
 [sdq]   prefetch=off  fadvise_dontneed=1
 
-[ Prompt: 5.9 t/s | Generation: 3.5 t/s ]
+[ Prompt: 9.6 t/s | Generation: 4.4 t/s ]
 
 
 Exiting...
-RUNMEM {"peak_kb": 4950432, "peak_gib": 4.721, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 7.78, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-fresh/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "4", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
+RUNMEM {"peak_kb": 4996996, "peak_gib": 4.766, "limit_gib": 8.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 6.25, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=3072", "SDQ_PRELOAD=0", "/tmp/opencode/flint-work/llama.cpp/build/bin/llama-cli", "-m", "/tmp/opencode/flint-work/llama.cpp/models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena18g_run1.log b/results/logs/v1/perf_c512_v1_arena18g_run1.log
new file mode 100644
index 0000000000000000000000000000000000000000..484504ccf9a175715e7053a0a23a3e0964550774
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena18g_run1.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.06 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 12.0 t/s | Generation: 5.5 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 20755568, "peak_gib": 19.794, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 19.15, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena18g_run2.log b/results/logs/v1/perf_c512_v1_arena18g_run2.log
new file mode 100644
index 0000000000000000000000000000000000000000..c5143f55376222426954f5448c916e205461aa30
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena18g_run2.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.06 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 13.0 t/s | Generation: 5.6 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 20755352, "peak_gib": 19.794, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 19.54, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena18g_run3.log b/results/logs/v1/perf_c512_v1_arena18g_run3.log
new file mode 100644
index 0000000000000000000000000000000000000000..0ebd2f1e351b7d4ce07090aa2088bdb9345bedac
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena18g_run3.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.06 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=7561  arena=18.00 GiB  (RAM 實測 RSS=18.17 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 7561
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 13.0 t/s | Generation: 5.7 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 20755712, "peak_gib": 19.794, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 21.96, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=18432", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena28g_run1.log b/results/logs/v1/perf_c512_v1_arena28g_run1.log
new file mode 100644
index 0000000000000000000000000000000000000000..f5923bf40a0e2a5ad5f3b3b6d2fddc18d0980339
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena28g_run1.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 11.5 t/s | Generation: 5.5 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31988764, "peak_gib": 30.507, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 29.95, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena28g_run2.log b/results/logs/v1/perf_c512_v1_arena28g_run2.log
new file mode 100644
index 0000000000000000000000000000000000000000..139846735b22430fe1da0a052ad42dc932c4fec7
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena28g_run2.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 12.6 t/s | Generation: 5.5 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31988676, "peak_gib": 30.507, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 29.03, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v1/perf_c512_v1_arena28g_run3.log b/results/logs/v1/perf_c512_v1_arena28g_run3.log
new file mode 100644
index 0000000000000000000000000000000000000000..6c40f28c580c2db70386ab8dfd7163c3e7de23db
--- /dev/null
+++ b/results/logs/v1/perf_c512_v1_arena28g_run3.log
@@ -0,0 +1,59 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 12.3 t/s | Generation: 6.2 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31988828, "peak_gib": 30.507, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 27.87, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=0", "SDQ_ROW_PARALLEL=0", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/ppl_epq.log b/results/logs/v1/ppl_prefill_ub64_t8.log
similarity index 100%
rename from results/logs/ppl_epq.log
rename to results/logs/v1/ppl_prefill_ub64_t8.log
diff --git a/results/logs/v2/greedy_t16.txt b/results/logs/v2/greedy_t16.txt
new file mode 100644
index 0000000000000000000000000000000000000000..84c4fd2f82b06d02f7c64b81cd118201ac404e24
--- /dev/null
+++ b/results/logs/v2/greedy_t16.txt
@@ -0,0 +1,2 @@
+ Paris is a true statement. The capital of France is London is a false. The capital of France is Berlin is a
+
diff --git a/results/logs/v2/greedy_t8.txt b/results/logs/v2/greedy_t8.txt
new file mode 100644
index 0000000000000000000000000000000000000000..f5e3f96f9a1ce6be6680de660e7cff069d078c32
--- /dev/null
+++ b/results/logs/v2/greedy_t8.txt
@@ -0,0 +1,2 @@
+ Paris is a simple true/false question. The answer is true. The capital of France is indeed Paris. The user is
+
diff --git a/results/logs/v2/kl_decode_ub1_t16_vs_native.log b/results/logs/v2/kl_decode_ub1_t16_vs_native.log
new file mode 100644
index 0000000000000000000000000000000000000000..842eafbfc65320e4247a6fd5b5dca04a3b247ae9
--- /dev/null
+++ b/results/logs/v2/kl_decode_ub1_t16_vs_native.log
@@ -0,0 +1,70 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.76 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+0.30.239.333 I cmn          init: llama threadpool init, n_threads = 16
+0.30.433.929 I 
+0.30.434.050 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.30.450.827 I kl_divergence: computing over 2 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.43.870.456 I kl_divergence: 13.42 seconds per pass - ETA 0.43 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1      13.7740 ±    4.6245       0.02254 ±    0.03440       0.03177 ±    0.01493     4.997 ±  2.174 %    95.276 ±  1.890 %
+   2      14.8027 ±    3.1804       0.01283 ±    0.02276       0.04406 ±    0.01523     4.051 ±  1.350 %    92.520 ±  1.654 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  14.802734 ±   3.180425
+Mean PPL(base)                :  14.614086 ±   3.091134
+Cor(ln(PPL(Q)), ln(PPL(base))):  99.44%
+Mean ln(PPL(Q)/PPL(base))     :   0.012826 ±   0.022762
+Mean PPL(Q)/PPL(base)         :   1.012909 ±   0.023056
+Mean PPL(Q)-PPL(base)         :   0.188648 ±   0.342994
+
+====== KL divergence statistics ======
+Mean    KLD:   0.044057 ±   0.015234
+Maximum KLD:   3.161583
+99.9%   KLD:   2.810790
+99.0%   KLD:   0.854633
+95.0%   KLD:   0.086102
+90.0%   KLD:   0.048945
+Median  KLD:   0.005596
+10.0%   KLD:   0.000030
+ 5.0%   KLD:   0.000006
+ 1.0%   KLD:   0.000000
+ 0.1%   KLD:  -0.000003
+Minimum KLD:  -0.000004
+
+====== Token probability statistics ======
+Mean    Δp: -0.309 ± 0.254 %
+Maximum Δp: 12.097%
+99.9%   Δp: 11.854%
+99.0%   Δp:  7.841%
+95.0%   Δp:  2.800%
+90.0%   Δp:  1.211%
+75.0%   Δp:  0.182%
+Median  Δp:  0.000%
+25.0%   Δp: -0.223%
+10.0%   Δp: -2.155%
+ 5.0%   Δp: -3.443%
+ 1.0%   Δp: -9.554%
+ 0.1%   Δp: -43.359%
+Minimum Δp: -52.532%
+RMS Δp    :  4.051 ± 1.350 %
+Same top p: 92.520 ± 1.654 %
+
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.82 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
diff --git a/results/logs/v2/kl_decode_ub1_t16_vs_v1.log b/results/logs/v2/kl_decode_ub1_t16_vs_v1.log
new file mode 100644
index 0000000000000000000000000000000000000000..9061bfddcd508af6ebfd6e2d476ff9ec99a5fb26
--- /dev/null
+++ b/results/logs/v2/kl_decode_ub1_t16_vs_v1.log
@@ -0,0 +1,70 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.76 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+0.30.646.682 I cmn          init: llama threadpool init, n_threads = 16
+0.30.827.678 I 
+0.30.827.803 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.30.844.516 I kl_divergence: computing over 2 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.46.567.395 I kl_divergence: 15.72 seconds per pass - ETA 0.52 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1      13.7740 ±    4.6245       0.02395 ±    0.02233       0.00000 ±    0.00000     0.001 ±  0.000 %    100.000 ±  0.000 %
+   2      14.8027 ±    3.1804       0.01198 ±    0.01117       0.00000 ±    0.00000     0.000 ±  0.000 %    100.000 ±  0.000 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  14.802734 ±   3.180425
+Mean PPL(base)                :  14.626524 ±   3.094576
+Cor(ln(PPL(Q)), ln(PPL(base))):  99.87%
+Mean ln(PPL(Q)/PPL(base))     :   0.011975 ±   0.011170
+Mean PPL(Q)/PPL(base)         :   1.012047 ±   0.011305
+Mean PPL(Q)-PPL(base)         :   0.176210 ±   0.179040
+
+====== KL divergence statistics ======
+Mean    KLD:   0.000001 ±   0.000001
+Maximum KLD:   0.000038
+99.9%   KLD:   0.000037
+99.0%   KLD:   0.000034
+95.0%   KLD:   0.000022
+90.0%   KLD:   0.000015
+Median  KLD:   0.000000
+10.0%   KLD:  -0.000011
+ 5.0%   KLD:  -0.000020
+ 1.0%   KLD:  -0.000034
+ 0.1%   KLD:  -0.000039
+Minimum KLD:  -0.000040
+
+====== Token probability statistics ======
+Mean    Δp: -0.000 ± 0.000 %
+Maximum Δp:  0.003%
+99.9%   Δp:  0.002%
+99.0%   Δp:  0.001%
+95.0%   Δp:  0.001%
+90.0%   Δp:  0.000%
+75.0%   Δp:  0.000%
+Median  Δp: -0.000%
+25.0%   Δp: -0.000%
+10.0%   Δp: -0.000%
+ 5.0%   Δp: -0.001%
+ 1.0%   Δp: -0.002%
+ 0.1%   Δp: -0.003%
+Minimum Δp: -0.004%
+RMS Δp    :  0.000 ± 0.000 %
+Same top p: 100.000 ± 0.000 %
+
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.82 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
diff --git a/results/logs/v2/kl_prefill_ub64_t8_vs_native.log b/results/logs/v2/kl_prefill_ub64_t8_vs_native.log
new file mode 100644
index 0000000000000000000000000000000000000000..b942f00f227a988d1519edb826e969724b4552c3
--- /dev/null
+++ b/results/logs/v2/kl_prefill_ub64_t8_vs_native.log
@@ -0,0 +1,72 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.76 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+0.27.947.775 I cmn          init: llama threadpool init, n_threads = 8
+0.28.110.989 I 
+0.28.111.120 I system_info: n_threads = 8 (n_threads_batch = 8) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.28.130.995 I kl_divergence: computing over 4 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.34.415.654 I kl_divergence: 6.28 seconds per pass - ETA 0.42 minutes
+
+chunk             PPL               ln(PPL(Q)/PPL(base))          KL Divergence              Δp RMS            Same top p
+   1      13.4089 ±    4.4805      -0.00934 ±    0.04792       0.02043 ±    0.00445     2.355 ±  0.333 %    92.913 ±  2.286 %
+   2      14.5720 ±    3.1356      -0.00591 ±    0.03286       0.02463 ±    0.00364     3.396 ±  0.658 %    90.157 ±  1.873 %
+   3       9.0785 ±    1.5308       0.00134 ±    0.02267       0.01938 ±    0.00251     3.090 ±  0.488 %    91.864 ±  1.402 %
+   4      10.8014 ±    1.6076       0.00923 ±    0.01822       0.01912 ±    0.00198     2.903 ±  0.398 %    92.323 ±  1.182 %
+
+====== Perplexity statistics ======
+Mean PPL(Q)                   :  10.801407 ±   1.607644
+Mean PPL(base)                :  10.702135 ±   1.575857
+Cor(ln(PPL(Q)), ln(PPL(base))):  99.25%
+Mean ln(PPL(Q)/PPL(base))     :   0.009233 ±   0.018216
+Mean PPL(Q)/PPL(base)         :   1.009276 ±   0.018385
+Mean PPL(Q)-PPL(base)         :   0.099271 ±   0.197680
+
+====== KL divergence statistics ======
+Mean    KLD:   0.019120 ±   0.001984
+Maximum KLD:   0.455408
+99.9%   KLD:   0.412640
+99.0%   KLD:   0.236318
+95.0%   KLD:   0.072383
+90.0%   KLD:   0.045335
+Median  KLD:   0.004286
+10.0%   KLD:   0.000022
+ 5.0%   KLD:   0.000003
+ 1.0%   KLD:  -0.000000
+ 0.1%   KLD:  -0.000001
+Minimum KLD:  -0.000001
+
+====== Token probability statistics ======
+Mean    Δp:  0.088 ± 0.129 %
+Maximum Δp: 28.966%
+99.9%   Δp: 21.585%
+99.0%   Δp:  7.792%
+95.0%   Δp:  3.637%
+90.0%   Δp:  1.888%
+75.0%   Δp:  0.336%
+Median  Δp:  0.000%
+25.0%   Δp: -0.142%
+10.0%   Δp: -1.699%
+ 5.0%   Δp: -3.238%
+ 1.0%   Δp: -7.473%
+ 0.1%   Δp: -20.033%
+Minimum Δp: -26.901%
+RMS Δp    :  2.903 ± 0.398 %
+Same top p: 92.323 ± 1.182 %
+
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=29.01 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
diff --git a/results/logs/v2/perf_c512_v2_arena28g_run1.log b/results/logs/v2/perf_c512_v2_arena28g_run1.log
new file mode 100644
index 0000000000000000000000000000000000000000..822b497b6651f0686351750ecd53f46003ab7927
--- /dev/null
+++ b/results/logs/v2/perf_c512_v2_arena28g_run1.log
@@ -0,0 +1,61 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.87 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 31.1 t/s | Generation: 14.5 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31980088, "peak_gib": 30.499, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 45.04, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v2/perf_c512_v2_arena28g_run2.log b/results/logs/v2/perf_c512_v2_arena28g_run2.log
new file mode 100644
index 0000000000000000000000000000000000000000..920a169cdf1f1db77d0bcbb900e4c575a679a8e3
--- /dev/null
+++ b/results/logs/v2/perf_c512_v2_arena28g_run2.log
@@ -0,0 +1,61 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.87 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 32.4 t/s | Generation: 16.2 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31980008, "peak_gib": 30.499, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 32.44, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v2/perf_c512_v2_arena28g_run3.log b/results/logs/v2/perf_c512_v2_arena28g_run3.log
new file mode 100644
index 0000000000000000000000000000000000000000..aacc8f0dbd78553fe355c13d4858499faba88ad3
--- /dev/null
+++ b/results/logs/v2/perf_c512_v2_arena28g_run3.log
@@ -0,0 +1,61 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one-sentence explanation of why the sky is blue.
+
+This is a classic physics question about Rayleigh scattering.
+
+The key points:
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.87 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 40.0 t/s | Generation: 17.3 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31979436, "peak_gib": 30.498, "limit_gib": null, "limit_exceeded": false, "swap_kb": 0, "wall_s": 30.47, "exit": 0, "cmd": ["env", "SDQ_PREFETCH=0", "SDQ_PAGER=1", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "512", "-b", "128", "-ub", "32", "-p", "Explain in one sentence why the sky is blue.", "-n", "32", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v2/ppl_prefill_ub64_t8.log b/results/logs/v2/ppl_prefill_ub64_t8.log
new file mode 100644
index 0000000000000000000000000000000000000000..78d5b272b664ac7c173baf77374ce62ec9f8c455
--- /dev/null
+++ b/results/logs/v2/ppl_prefill_ub64_t8.log
@@ -0,0 +1,30 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.76 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+0.27.382.921 I cmn          init: llama threadpool init, n_threads = 8
+0.27.578.587 I 
+0.27.578.686 I system_info: n_threads = 8 (n_threads_batch = 8) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 
+0.27.578.703 I perplexity: tokenizing the input ..
+0.27.585.073 I perplexity: tokenization took 6.365 ms
+0.27.585.087 I perplexity: calculating perplexity over 4 chunks, n_ctx=256, batch_size=256, n_seq=1
+0.33.385.452 I perplexity: 5.80 seconds per pass - ETA 0.38 minutes
+[1]13.4089,[2]14.5720,[3]9.0785,[4]10.8014,
+0.50.490.263 I Final estimate: PPL = 10.8014 +/- 1.60764
+
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=29.01 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
diff --git a/results/logs/v2/v2_32g_run1.log b/results/logs/v2/v2_32g_run1.log
new file mode 100644
index 0000000000000000000000000000000000000000..672f41e6a4b8ef051aba2d80e505430a6b7695db
--- /dev/null
+++ b/results/logs/v2/v2_32g_run1.log
@@ -0,0 +1,57 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 33.7 t/s | Generation: 17.3 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31953556, "peak_gib": 30.473, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 30.09, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v2/v2_32g_run2.log b/results/logs/v2/v2_32g_run2.log
new file mode 100644
index 0000000000000000000000000000000000000000..9852be4662ff6363da8620a0f623ad8f2bb29690
--- /dev/null
+++ b/results/logs/v2/v2_32g_run2.log
@@ -0,0 +1,57 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 30.8 t/s | Generation: 13.7 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31953620, "peak_gib": 30.473, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 30.2, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/logs/v2/v2_32g_run3.log b/results/logs/v2/v2_32g_run3.log
new file mode 100644
index 0000000000000000000000000000000000000000..80be63f0dac2719ba9306436097fe545d7133c29
--- /dev/null
+++ b/results/logs/v2/v2_32g_run3.log
@@ -0,0 +1,57 @@
+[sdq] MoE 層 47 / 共 49 層（dense lead 2）：expert 權重合計 21.47 GiB、12032 個 expert
+[sdq] 統一槽位 2555904 B（最大 expert 大小；平均 1915708 B）
+[sdq] 使用 O_DIRECT 讀取模型檔（1 個 shard，不佔用 kernel page cache）
+[sdq] I/O 段平行：24 條讀取執行緒 × 1096 KB 對齊緩衝（gate/up/down 同時讀，每段再切 1 塊）
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.77 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+[sdq] 已預載全部 expert（12061 slots）→ 之後 row-parallel 可用
+
+
+Loading model... 
+
+▄▄ ▄▄
+██ ██
+██ ██  ▀▀█▄ ███▄███▄  ▀▀█▄    ▄████ ████▄ ████▄
+██ ██ ▄█▀██ ██ ██ ██ ▄█▀██    ██    ██ ██ ██ ██
+██ ██ ▀█▄██ ██ ██ ██ ▀█▄██ ██ ▀████ ████▀ ████▀
+                                    ██    ██
+                                    ▀▀    ▀▀
+
+build      : b1-970b51c
+model      : models/flint-UD-Q4_K_XL.gguf
+ftype      : Q4_K - Medium
+modalities : text
+
+available commands:
+  /exit or Ctrl+C     stop or exit
+  /regen              regenerate the last response
+  /clear              clear the chat history
+  /read <file>        add a text file
+  /glob <pattern>     add text files using globbing pattern
+
+
+
+> Explain in one sentence why the sky is blue.
+
+[Start thinking]
+
+The user is asking for a one
+[sdq]   page-map: n_layer=49 n_expert=256 n_expert_used=6 n_embd=2048 n_ff_exp=512 | L2E0 gate=986105536/450560 up=1104733888/450560 down=799327936/720896 (off/size) ne0=2048 ne1=512 d_ne0=512 d_ne1=2048
+[sdq] expert pager ready: models/flint-UD-Q4_K_XL.gguf
+[sdq]   熱門 expert 釘選：0 個（無剖面，用動態 LRU/LFU）
+[sdq]   slot=2.44 MiB (2555904 B)  slots=12061  arena=28.71 GiB  (RAM 實測 RSS=28.88 GiB)
+[sdq]   evict_policy=lru  stale_window=640
+[sdq]   槽位：單一尺寸 2.44 MiB × 12061
+[sdq]   prefetch=off  fadvise_dontneed=1
+
+[ Prompt: 34.5 t/s | Generation: 13.8 t/s ]
+
+
+Exiting...
+RUNMEM {"peak_kb": 31953600, "peak_gib": 30.473, "limit_gib": 32.0, "limit_exceeded": false, "swap_kb": 0, "wall_s": 30.34, "exit": 0, "cmd": ["env", "SDQ_PAGER=1", "SDQ_PREFETCH=0", "SDQ_ARENA_MB=29400", "SDQ_PRELOAD=1", "SDQ_ROW_PARALLEL=1", "build/bin/llama-cli", "-m", "models/flint-UD-Q4_K_XL.gguf", "-c", "128", "-b", "32", "-ub", "16", "-p", "Explain in one sentence why the sky is blue.", "-n", "8", "-t", "16", "--temp", "0", "-st", "--simple-io"], "method": "procfs-rss-watchdog"}
diff --git a/results/test-status.md b/results/test-status.md
index 70aa9b3c94deb30783570d041d938546aa7f7e42..c0c5c94c631189f86184928f51611e9e6c4bc589 100644
--- a/results/test-status.md
+++ b/results/test-status.md
@@ -1,47 +1,63 @@
 # Validation status
 
-Re-verified on 2026-10-23 after a correctness audit. Environment: Intel Xeon Platinum 8559C, container quota 16 CPUs (benchmarks used 4 threads), cgroup v2 `memory.max` 104,000,000,000 bytes (~96.9 GiB), `memory.swap.max` 2,936,848,384 bytes (~2.74 GiB), Amazon Linux 2023, kernel 6.18.48, CPU only. The container's cgroup v2 mount is read-only and there is no systemd manager, so no cgroup-enforced memory limit could be applied.
+Environment: Intel Xeon Platinum 8559C, container quota 16 CPUs, cgroup v2 `memory.max` 104,000,000,000 bytes (~96.9 GiB), `memory.swap.max` 2,936,848,384 bytes (~2.74 GiB), Amazon Linux 2023, kernel 6.18.48, CPU only. The container's cgroup v2 mount is read-only and there is no systemd manager, so no cgroup-enforced memory limit could be applied.
 
-llama.cpp base: `970b51ce7c9ba5fff0726c5e9994f1c2f0e4eedc` plus `flint-epq.patch`.
+llama.cpp base: `970b51ce7c9ba5fff0726c5e9994f1c2f0e4eedc` plus `flint-epq.patch` (sha256 `15c6554271d0c77357d96e628546f8ba624e6834c4c90c69615c90c2ac2310ea`).
+
+Versions: **v1** = expert paging, mode C (`SDQ_PRELOAD=0 SDQ_ROW_PARALLEL=0`). **v2** = v1 + preload + row-parallel for gated experts (`SDQ_PRELOAD=1 SDQ_ROW_PARALLEL=1`, arena ≥ 28.6 GiB). **native** = `SDQ_PAGER=0`.
 
 ## Checks
 
-| Check | Result |
-|---|---|
-| Clean checkout of `970b51c` + `flint-epq.patch` applies; CPU build of `llama-cli`, `llama-server`, `llama-completion`, `llama-perplexity` | PASS (fresh worktree, configure command from README) |
-| Flint GGUF header / expert map | PASS: 49 layers (2 dense + 47 MoE), 256 experts, top-6, 12,032 routed experts, 21.47 GiB routed expert weights |
-| Expert type coverage (`IQ3_S`, `IQ4_XS`, `Q4_K`, `Q5_K`, `Q6_K`, `Q8_0`) | PASS: all types have dot-product kernels; gate and up share one activation type in all 47 layers |
-| Smoke test (README command, `-c 2048`, arena 512 MiB, `--single-turn`) | PASS: exit 0, answer produced, peak process-tree RSS 2.392 GiB, swap 0 |
-| Greedy output, 24 tokens, EPQ vs native | PASS: byte-identical (fresh build) |
-| Perplexity, 4 × 256-token chunks: native / EPQ | 10.7790 / 10.8014 (+0.21 %) |
-| KL divergence vs native, EPQ | mean 0.019120 ± 0.001984; median 0.004286; max 0.455; top-1 agreement 92.32 % |
-| Control: native ubatch 16 vs 64 | mean KLD 0.024005 ± 0.003389; max 1.136; top-1 agreement 93.11 % |
-| Layer-wise MoE output vs native, layers 2–6 (same input) | PASS: max relative difference ≤ 1.6e-5 |
-| Memory tiers 2 / 4 / 8 / 16 / 32 GiB (RSS watchdog, 3 runs each) | PASS: all 15 runs exit 0, swap 0, never over threshold; peaks 1.970 / 2.721 / 4.721 / 11.721 / 19.720 GiB |
-| cgroup-v2 `MemoryMax` runs (`tools/runmem_cgroup.py`) | NOT RUN: no writable cgroup v2 / systemd here. The script is untested |
-
-## Defects found and fixed in this pass
-
-1. **Bad debug hook (debug path only):** `sdq_maybe_dump_tensor` (enabled by `SDQ_MOE_OUT_DUMP`) wrote the tensor to a file but never copied it into its output. Downstream nodes read uninitialized memory whenever the dump was enabled, which made early layer comparisons look wrong. Fixed: the dump node now copies its input. The production path (no dump env) never used this node and was unaffected.
-2. **SwiGLU clamp mismatch:** the pager clamped `silu(gate)` to `[-limit, limit]`; upstream `build_moe_ffn` clamps only the upper bound (`ggml_clamp(-INF, limit)`). Fixed to match upstream. This checkpoint has no `swiglu_clamp` metadata, so the clamp is inactive here and the numbers above are unaffected.
-3. **Published patch was stale:** the patch on Hugging Face predated fixes 1 and 2. Re-generated from the verified tree and re-uploaded.
-4. **Smoke command in the model card was not the one that was run** (missing `--simple-io`, different context). Model card now shows the command that was executed.
-5. **`tools/runmem_cgroup.py` and the cgroup claim were presented as usable.** Now labelled untested.
-6. **Earlier tier throughput was from single runs** and varied by ~±15 % between runs. Replaced with medians of 3 runs.
-
-## EPQ version coverage
-
-- v1 (expert paging, mode C): active in all runs.
-- v2 preload: code present, requires an arena of at least 28.6 GiB; never triggered in these tests (the 24 GiB run had 10,082 slots, below the 12,032 needed).
-- v2 row-parallel: Nemotron no-gate only; inactive for Flint.
+| Check | Version | Result |
+|---|---|---|
+| Clean checkout of `970b51c` + patch applies; CPU build | v1, v2 | PASS (fresh worktree, configure command from README) |
+| Flint GGUF header / expert map | — | PASS: 49 layers (2 dense + 47 MoE), 256 experts, top-6, 12,032 routed experts, 21.47 GiB routed expert weights |
+| Expert type coverage (`IQ3_S`, `IQ4_XS`, `Q4_K`, `Q5_K`, `Q6_K`, `Q8_0`) | — | PASS: all types have dot-product kernels; gate and up share one activation type in all 47 layers |
+| Smoke test (README command, `-c 2048`, arena 512 MiB, `--single-turn`) | v1 | PASS: exit 0, answer produced, peak process-tree RSS 2.392 GiB, swap 0 |
+| Greedy, 24 tokens, 8 threads | v1, v2 | PASS: identical to native |
+| Greedy, 24 tokens, 16 threads | v1, v2 | v1 = v2; both differ from native at char 13 (near-tie argmax flip) |
+| Prefill perplexity, `-ub 64`, 8 threads, 4 chunks | native / v1 / v2 | 10.7790 / 10.8014 / 10.8014 |
+| Prefill KL vs native | v1, v2 | mean 0.019120 ± 0.001984; median 0.004286; max 0.455; top-1 92.32 % (identical for v1 and v2) |
+| Decode KL vs native, `-ub 1`, 16 threads, 2 chunks | v1 | mean 0.044057; max 3.16; top-1 92.52 % |
+| Decode KL vs native, `-ub 1`, 16 threads, 2 chunks | v2 | mean 0.044057; max 3.16; top-1 92.52 % |
+| Decode v2 vs v1, `-ub 1`, 16 threads | v2 vs v1 | mean KLD 1e-6 (max 3.8e-5); top-1 100 %. Row-parallel is numerically equivalent to mode C but not bit-identical |
+| Control: native `-ub 64` vs native `-ub 1`, 16 threads | native | mean KLD 0.034464 ± 0.008439; top-1 90.16 % |
+| Control: native `-ub 16` vs `-ub 64`, prefill | native | mean KLD 0.024005 ± 0.003389; top-1 93.11 % |
+| Layer-wise MoE output vs native, layers 2–6 (same input) | v1 | max relative difference ≤ 1.6e-5 |
+| Row-parallel active check (`SDQ_DBG_OP`) | v2 | `rowpar=1` on decode ops with 6 needed experts at 16 threads; `rowpar=0` with `SDQ_ROW_PARALLEL=0` |
+| Preload check | v2 | "已預載全部 expert（12061 slots）" with a 28.7 GiB arena |
+| Throughput, context 512, batch 128, ubatch 32, 32 tokens, 16 threads (3 runs, median) | native / v1 / v2 | Generation 21.5 / 5.6 / 16.2 t/s; prompt 54.4 / 13.0 / 32.4 t/s. Peak RSS 24.3 / 19.8 / 30.5 GiB |
+| Memory tiers 2 / 4 / 8 / 16 / 32 GiB, 16 threads (3 runs each) | v1 | PASS: all 15 runs exit 0, swap 0, never over threshold; peaks 2.014 / 2.765 / 4.766 / 11.767 / 19.766 GiB |
+| 32 GiB tier, 28.7 GiB arena, 16 threads (3 runs) | v2 | PASS: peak 30.473 GiB, swap 0, never over threshold |
+| cgroup-v2 `MemoryMax` runs (`tools/runmem_cgroup.py`) | — | NOT RUN: no writable cgroup v2 / systemd here. The script is untested |
+
+## Defects found and fixed
+
+1. **Bad debug hook (debug path only):** `sdq_maybe_dump_tensor` (enabled by `SDQ_MOE_OUT_DUMP`) wrote the tensor to a file but never copied it into its output, so downstream nodes read uninitialized memory. Fixed; the production path never used this node.
+2. **SwiGLU clamp mismatch:** the pager clamped `silu(gate)` to `[-limit, limit]`; upstream clamps only the upper bound. Fixed to match upstream.
+3. **`swiglu_limit` was dropped (found in this pass):** `sdq_build_moe` ignored its `swiglu_limit` argument and the op always used 0. Fixed; the value now reaches the op. This checkpoint has limit 0, so the numbers are unaffected.
+4. **Published patch was stale:** the patch on Hugging Face predated these fixes. Re-generated from the verified tree and re-uploaded.
+5. **Smoke command in the model card was not the one that was run.** The card now shows the executed command.
+6. **`tools/runmem_cgroup.py` and the cgroup claim were presented as usable.** Now labeled untested.
+7. **Earlier memory-tier throughput came from single runs** and varied by ~±15 %. Replaced with medians of 3 runs.
+8. **Earlier tier tables used 4 threads; `bench.sh` defaults to 16.** The tier tables are now labeled with 16 threads; the earlier 4-thread numbers are superseded.
+9. **Row-parallel was inactive for Flint.** Added the gated (gate/up/down) path so v2 applies to Flint; verified active and numerically equivalent to v1 (see checks).
 
 ## Known limits and open items
 
-- EPQ output is not bit-identical to native llama.cpp. Differences come from float summation order and are amplified by 8-bit activation quantization; native llama.cpp itself shows the same magnitude of KLD under a ubatch change (control above). No functional mismatch was found.
-- The memory tiers use an RSS watchdog that samples every 20 ms and counts file-backed pages as reported by `/proc`. They are not kernel-enforced limits. A true cgroup-enforced test must be run on a host with a writable cgroup v2 hierarchy and systemd (`tools/runmem_cgroup.py`, untested).
-- Benchmarks set `SDQ_PRELOAD=0`, so the preload / row-parallel path is not exercised by the tier numbers. The 18 GiB arena at the 32 GiB tier holds 7,561 of 12,032 experts, so eviction occurs.
-- The accuracy tests use about 1,000 scored tokens from one text sample. They do not cover long-context behavior or other prompts beyond the greedy check.
+- EPQ output is not bit-identical to native llama.cpp. Float summation order differs, and 8-bit activation quantization amplifies the difference through the layers. Native itself shows a KL divergence of the same magnitude under a ubatch change. No functional mismatch was found.
+- Greedy identity is fragile: at 16 threads both v1 and v2 diverge from native after 12 characters.
+- The memory tiers use an RSS watchdog that samples every 20 ms and counts file-backed pages as reported by `/proc`. They are not kernel-enforced limits. A true cgroup-enforced test requires a host with a writable cgroup v2 hierarchy and systemd (`tools/runmem_cgroup.py`, untested).
+- v2 only runs in the 32 GiB tier, where the arena can hold all experts. The 2–16 GiB tiers are v1 only.
+- Accuracy tests use about 1,000 scored tokens from one text sample. They do not cover long-context behavior or other prompts beyond the greedy checks.
+- The v2 prompt-phase speedup is likely due to preload removing SSD reads during prefill; this was not isolated by a separate measurement.
 
 ## Raw logs
 
-`results/logs/`: KL and perplexity logs (`kl_epq_vs_native.log`, `kl_native_ub16_vs_native.log`, `ppl_kl_native_base_ub64.log`, `ppl_epq.log`), greedy outputs (`greedy_native.txt`, `greedy_epq_fresh.txt`), the text corpus (`ppl_corpus.txt`), and per-run memory-tier logs with the 3-run summary (`tiers_3runs.txt`).
+`results/logs/{native,v1,v2}/` hold the per-version logs; `results/logs/common/` holds the corpus and summaries:
+
+- Prefill accuracy: `v1/ppl_prefill_ub64_t8.log`, `v1/kl_prefill_ub64_t8_vs_native.log`, `v2/ppl_prefill_ub64_t8.log`, `v2/kl_prefill_ub64_t8_vs_native.log`, `native/ppl_kl_base_prefill_ub64_t8.log`, `native/kl_ub16_vs_native_ub64.log`
+- Decode accuracy: `v1/kl_decode_ub1_t16_vs_native.log`, `v2/kl_decode_ub1_t16_vs_native.log`, `v2/kl_decode_ub1_t16_vs_v1.log`, `native/ppl_kl_base_decode_ub1_t16.log`, `native/kl_ub64_vs_native_ub1_t16.log`
+- Greedy: `native/greedy_t8.txt`, `native/greedy_t16.txt`, `v1/greedy_t8.txt`, `v1/greedy_t16.txt`, `v2/greedy_t8.txt`, `v2/greedy_t16.txt`
+- Throughput: `*/perf_c512_*.log`, `common/perf_c512_summary.txt`
+- Memory: `v1/mem_v1_*.log`, `v2/v2_32g_run*.log`, `common/mem_summary.txt`
diff --git a/src/sdq_moe.cpp b/src/sdq_moe.cpp
index 38ad23572e527bed39de1bd98472de6eace3c490..14cc23e500e9f23c352d2ce479c10312c5d4207e 100644
--- a/src/sdq_moe.cpp
+++ b/src/sdq_moe.cpp
@@ -329,6 +329,19 @@ static size_t env_partial_cap_mb() {
 
 // ---------------------------------------------------------------- 單一 expert 的數學
 // 從 arena 的槽位記憶體算出一個 expert 對多個 token 的輸出（加權後直接累加到 out）
+// SwiGLU 啟動，與上游 build_moe_ffn 一致（見 llama-graph.cpp 的 LLM_FFN_SILU 分支）：
+// limit > eps 時 up 兩側限制、silu(gate) 只限制上界；否則 silu(g)*u。
+// expert_forward 與 row-parallel 共用，確保兩條路徑數學完全相同。
+static inline float gated_act(float g, float u, float swiglu_limit) {
+    if (swiglu_limit > 1e-6f) {
+        const float uc = std::max(-swiglu_limit, std::min(swiglu_limit, u));
+        float gc = g / (1.0f + std::exp(-g));
+        gc = std::min(swiglu_limit, gc);
+        return gc * uc;
+    }
+    return (g / (1.0f + std::exp(-g))) * u;
+}
+
 static void expert_forward(const expert_desc & ed, const uint8_t * base,
                            const float * cur, int64_t n_embd,
                            const std::vector<std::pair<int32_t, int32_t>> & toks,
@@ -405,19 +418,7 @@ static void expert_forward(const expert_desc & ed, const uint8_t * base,
 
         PROF_T0_x(_td);   // SiLU + 加權求和
         for (int64_t r = 0; r < ed.down_ne0; r++) {
-            const float g = gate[(size_t) r];
-            const float u = upv[(size_t) r];
-            float act;
-            if (swiglu_limit > 1e-6f) {
-                const float uc = std::max(-swiglu_limit, std::min(swiglu_limit, u));
-                float gc = g / (1.0f + std::exp(-g));
-                // 與上游 build_moe_ffn 一致：silu(gate) 只限制上界（ggml_clamp(-INF, limit)）
-                gc = std::min(swiglu_limit, gc);
-                act = gc * uc;
-            } else {
-                act = (g / (1.0f + std::exp(-g))) * u;
-            }
-            h[(size_t) r] = act;
+            h[(size_t) r] = gated_act(gate[(size_t) r], upv[(size_t) r], swiglu_limit);
         }
         quantize_act(act_down, h.data(), hq, ed.down_ne0);
 
@@ -614,10 +615,13 @@ static void moe_row_parallel(ggml_tensor * dst, int ith, int nth, layer_rec * R,
     struct ed_info {
         const uint8_t * wu = nullptr;
         const uint8_t * wd = nullptr;
+        const uint8_t * wg = nullptr;   // gated 模型才有：gate 權重
         size_t up_row = 0;
         size_t dn_row = 0;
+        size_t g_row = 0;
         vec_dot_fn f_up = nullptr;
         vec_dot_fn f_down = nullptr;
+        vec_dot_fn f_gate = nullptr;
         const expert_desc * ed = nullptr;
         int64_t n_te = 0;
     };
@@ -633,9 +637,16 @@ static void moe_row_parallel(ggml_tensor * dst, int ith, int nth, layer_rec * R,
         q.dn_row = ggml_row_size(ed.type[part::DOWN], ed.down_ne0);
         q.f_up   = dot_for(ed.type[part::UP]);
         q.f_down = dot_for(ed.type[part::DOWN]);
+        if (!R->no_gate) {
+            q.g_row  = ggml_row_size(ed.type[part::GATE], ed.ne0);
+            q.f_gate = dot_for(ed.type[part::GATE]);
+        }
         if (sl >= 0) {
             q.wu = P.slot_ptr(sl) + ed.slot_off[part::UP];
             q.wd = P.slot_ptr(sl) + ed.slot_off[part::DOWN];
+            if (!R->no_gate) {
+                q.wg = P.slot_ptr(sl) + ed.slot_off[part::GATE];
+            }
         }
     }
 
@@ -655,6 +666,9 @@ static void moe_row_parallel(ggml_tensor * dst, int ith, int nth, layer_rec * R,
             if (!q.wu || !q.f_up) {
                 continue;
             }
+            if (!R->no_gate && (!q.wg || !q.f_gate)) {
+                continue;
+            }
             const auto & toks = R->tok_of_expert[(size_t) R->need[(size_t) j]];
             for (size_t ti = 0; ti < toks.size(); ti++) {
                 const int32_t t = toks[ti].first;
@@ -669,8 +683,17 @@ static void moe_row_parallel(ggml_tensor * dst, int ith, int nth, layer_rec * R,
                     float acc = 0.0f;
                     dot_row(q.f_up, q.wu + (size_t) r * q.up_row, xq,
                             q.up_row, (int) q.ed->ne0, acc);
-                    acc = acc > 0.0f ? acc : 0.0f;
-                    h[(size_t) r] = acc * acc;
+                    if (R->no_gate) {
+                        // Nemotron LatentMoE：relu(x)^2，沒有 gate
+                        acc = acc > 0.0f ? acc : 0.0f;
+                        h[(size_t) r] = acc * acc;
+                    } else {
+                        // gated（Flint）：silu(gate)*up，與 expert_forward 相同的 SwiGLU
+                        float g = 0.0f;
+                        dot_row(q.f_gate, q.wg + (size_t) r * q.g_row, xq,
+                                q.g_row, (int) q.ed->ne0, g);
+                        h[(size_t) r] = gated_act(g, acc, R->swiglu_limit);
+                    }
                 }
             }
         }
@@ -840,7 +863,14 @@ static void moe_custom_op(ggml_tensor * dst, int ith, int nth, void * userdata)
     if (ith == 0) {
         R->use_rowpar = false;
         R->rowpar_miss = 0;
-        if (rowpar_env && R->no_gate && n_need > 0 && n_need < nth) {
+        // gated（Flint）需要 gate 與 up 共用同一個 activation 量化型別，才能共用 xq。
+        bool types_ok = true;
+        if (!R->no_gate && n_need > 0) {
+            const expert_desc & e0 = L->experts[(size_t) R->need[0]];
+            types_ok = !e0.fused_gate_up && dot_for(e0.type[part::GATE]) && dot_for(e0.type[part::UP]) &&
+                       act_type_for(e0.type[part::GATE]) == act_type_for(e0.type[part::UP]);
+        }
+        if (rowpar_env && n_need > 0 && n_need < nth && types_ok) {
             int miss = 0;
             for (int i = 0; i < n_need; i++) {
                 if (!P.resident(R->il, R->need[(size_t) i])) {
@@ -1088,8 +1118,10 @@ ggml_tensor * sdq_build_moe_ex(ggml_cgraph * gf, ggml_context * ctx, ggml_tensor
 ggml_tensor * sdq_build_moe(ggml_cgraph * gf, ggml_context * ctx, ggml_tensor * cur,
                             ggml_tensor * ids, ggml_tensor * weights, int il, int64_t n_expert,
                             int64_t n_expert_used, float swiglu_limit) {
-    (void) swiglu_limit;
-    return sdq_build_moe_ex(gf, ctx, cur, ids, weights, il, n_expert, n_expert_used, false, false);
+    ggml_tensor * out = sdq_build_moe_ex(gf, ctx, cur, ids, weights, il, n_expert, n_expert_used, false, false);
+    // 把 swiglu limit 傳給 op（先前這裡 (void) 丟掉，造成 clamp 永遠不生效）
+    rec_for(il)->swiglu_limit = swiglu_limit;
+    return out;
 }
 
 } // namespace sdq