v5: 移除臨時診斷；A/B 驗證 512/256/128MB 全部 all_passed=true

patches/0001-st-expert-pager.patch

@@ -53,7 +53,7 @@ index 1112ad885..e335a9d5f 100644
  
      if (weight_before_ffn) {
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 1f3b80b08..662491537 100644
+index 1f3b80b08..c7b77476e 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
 @@ -1,5 +1,7 @@
@@ -73,22 +73,11 @@ index 1f3b80b08..662491537 100644
      const auto & split_mode   = params.split_mode;
      const bool use_mlock      = params.load_mode == LLAMA_LOAD_MODE_MLOCK || params.load_mode == LLAMA_LOAD_MODE_MMAP_MLOCK;
      const auto & tensor_split = params.tensor_split;
-@@ -1821,7 +1825,24 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -1821,7 +1825,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
      // per-tensor activation precision policy
      prec_policy.load(ml, *this);
  
 -    ml.init_mappings(true, use_mlock ? &pimpl->mlock_mmaps : nullptr);
-+    if (getenv("ST_VERBOSE")) {
-+        FILE * f = fopen("/proc/self/status", "r");
-+        char line[256];
-+        while (f && fgets(line, sizeof line, f)) {
-+            if (!strncmp(line, "VmRSS:", 6)) { fprintf(stderr, "[st] init_mappings 前 VmRSS%s", line + 6); break; }
-+        }
-+        if (f) fclose(f);
-+        fprintf(stderr, "  (prefetch=%d use_mmap=%d check_tensors=%d)\n",
-+                (int) st::will_page(arch_name().c_str()), (int) ml.use_mmap, (int) ml.check_tensors);
-+    }
-+
 +    // ST expert pager：開著分頁時**不能** prefetch。
 +    // llama.cpp 預設 prefetch=true 會用 MAP_POPULATE 把整個 GGUF（2.45 GiB）
 +    // 在「載入模型」那一瞬间全部 fault 進 page cache —— 那一刻的 RSS 峰值
@@ -99,24 +88,7 @@ index 1f3b80b08..662491537 100644
      pimpl->mappings.reserve(ml.mappings.size());
  
      // create the backend buffers
-@@ -1956,6 +1977,16 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
-         });
-     }
- 
-+    if (getenv("ST_VERBOSE")) {
-+        FILE * f = fopen("/proc/self/status", "r");
-+        char line[256];
-+        while (f && fgets(line, sizeof line, f)) {
-+            if (!strncmp(line, "VmRSS:", 6)) { fprintf(stderr, "[st] load_all_data 前 VmRSS%s", line + 6); break; }
-+        }
-+        if (f) fclose(f);
-+        fprintf(stderr, "\n");
-+    }
-+
-     // load tensor data
-     for (auto & [ctx, buf_map] : ctx_buf_maps) {
-         if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
-@@ -1969,6 +2000,48 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -1969,6 +1979,35 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
          }
      }
  
@@ -140,19 +112,6 @@ index 1f3b80b08..662491537 100644
 +            by_name.emplace(nm, t);
 +        }
 +
-+        if (getenv("ST_VERBOSE")) {
-+            FILE * f = fopen("/proc/self/status", "r");
-+            char line[256];
-+            while (f && fgets(line, sizeof line, f)) {
-+                if (!strncmp(line, "VmRSS:", 6)) {
-+                    fprintf(stderr, "[st] 模型載入完成後 VmRSS%s", line + 6);
-+                    break;
-+                }
-+            }
-+            if (f) fclose(f);
-+            fprintf(stderr, "\n");
-+        }
-+
 +    if (getenv("ST_VERBOSE") && !map_base) {
 +            LLAMA_LOG_INFO("%s: [st] 沒有 mapping（mappings=%zu files=%zu），無法分頁\n",
 +                           __func__, pimpl->mappings.size(), ml.files.size());

validate/ab-test.json

@@ -1,5 +1,5 @@
 {
-  "when": "2026-10-07T16:26:13+0200",
+  "when": "2026-10-07T16:27:36+0200",
   "model": "/root/work/models/SmallThinker-4B-A0.6B-Instruct.Q4_K.gguf",
   "model_size_mib": 2508,
   "prompt": "Hello, who are you?",
@@ -12,590 +12,582 @@
         "ST_PAGER": "0"
       },
       "rc": 0,
-      "seconds": 9.03,
+      "seconds": 8.77,
       "peak": {
-        "total_rss_gb": 2.5644,
+        "total_rss_gb": 2.5642,
         "anon_rss_gb": 0.1033,
-        "file_rss_gb": 2.5587,
+        "file_rss_gb": 2.5584,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 2.5644
+        "hwm_rss_gb": 2.5642
       },
       "decode": {
-        "total_rss_gb": 2.5644,
+        "total_rss_gb": 2.5642,
         "note": "prefill 後的峰值；prefill 會摸遍全部 expert，所以 peak 不代表 decode"
       },
       "rss_timeline_mib": [
         [
-          0.0,
-          0.6
+          0.001,
+          0.9
         ],
         [
           0.207,
-          72.8
+          75.7
         ],
         [
-          0.413,
-          71.3
+          0.412,
+          76.6
         ],
         [
           0.618,
-          117.7
+          131.8
         ],
         [
-          0.824,
-          187.4
+          0.825,
+          201.6
         ],
         [
-          1.03,
-          255.4
+          1.034,
+          272.6
         ],
         [
-          1.238,
-          326.7
+          1.243,
+          344.4
         ],
         [
           1.451,
-          396.2
-        ],
-        [
-          1.66,
-          467.0
-        ],
-        [
-          1.87,
-          537.0
+          415.4
         ],
         [
-          2.08,
-          608.5
+          1.661,
+          486.4
         ],
         [
-          2.291,
-          680.3
+          1.871,
+          558.2
         ],
         [
-          2.504,
-          752.1
+          2.082,
+          628.2
         ],
         [
-          2.723,
-          826.1
+          2.294,
+          700.2
         ],
         [
-          2.938,
-          899.4
+          2.507,
+          772.2
         ],
         [
-          3.154,
-          971.1
+          2.725,
+          845.8
         ],
         [
-          3.37,
-          1044.4
+          2.945,
+          919.8
         ],
         [
-          3.586,
-          1118.2
+          3.167,
+          995.3
         ],
         [
-          3.803,
-          1191.2
+          3.4,
+          1072.8
         ],
         [
-          4.022,
-          1264.2
+          3.618,
+          1147.3
         ],
         [
-          4.241,
-          1338.3
+          3.836,
+          1220.1
         ],
         [
-          4.462,
-          1412.8
+          4.056,
+          1293.9
         ],
         [
-          4.695,
-          1489.0
+          4.276,
+          1367.7
         ],
         [
-          4.924,
-          1565.8
+          4.509,
+          1444.9
         ],
         [
-          5.154,
-          1642.8
+          4.735,
+          1519.7
         ],
         [
-          5.382,
-          1718.1
+          4.966,
+          1597.7
         ],
         [
-          5.608,
-          1795.6
+          5.198,
+          1675.3
         ],
         [
-          5.835,
-          1871.4
+          5.438,
+          1754.3
         ],
         [
-          6.063,
-          1947.9
+          5.665,
+          1830.3
         ],
         [
-          6.291,
-          2024.5
+          5.894,
+          1907.3
         ],
         [
-          6.522,
-          2100.2
+          6.123,
+          1984.9
         ],
         [
-          6.755,
-          2177.5
+          6.357,
+          2061.1
         ],
         [
-          6.987,
-          2255.8
+          6.59,
+          2140.4
         ],
         [
-          7.221,
-          2334.3
+          6.821,
+          2218.9
         ],
         [
-          7.459,
-          2411.6
+          7.052,
+          2295.9
         ],
         [
-          7.696,
-          2487.6
+          7.284,
+          2372.7
         ],
         [
-          7.934,
-          2566.1
+          7.524,
+          2452.5
         ],
         [
-          8.173,
-          2620.2
+          7.782,
+          2537.0
         ],
         [
-          8.414,
-          2626.0
+          8.03,
+          2620.0
         ],
         [
-          8.665,
-          2626.0
+          8.276,
+          2625.7
         ],
         [
-          8.918,
-          2626.0
+          8.525,
+          2625.7
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 198.1 t/s | Generation: 55.6 t/s ]",
+      "timing_line": "[ Prompt: 190.1 t/s | Generation: 61.3 t/s ]",
       "output": "Hello! 😊 I'm DeepSeek-R1, your friendly AI assistant. I'm here to help with all kinds of questions, learning, and creative tasks",
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true
     },
-    "budget1024mb": {
+    "budget512mb": {
       "env": {
         "ST_PAGER": "1",
-        "ST_RAM_BUDGET_MB": "1024",
+        "ST_RAM_BUDGET_MB": "512",
         "ST_RESERVE_MB": "200"
       },
       "rc": 0,
-      "seconds": 7.6,
+      "seconds": 7.54,
       "peak": {
-        "total_rss_gb": 1.3011,
+        "total_rss_gb": 0.8332,
         "anon_rss_gb": 0.103,
-        "file_rss_gb": 1.3028,
+        "file_rss_gb": 0.8308,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 1.3011
+        "hwm_rss_gb": 0.8332
       },
       "decode": {
-        "total_rss_gb": 1.3011,
+        "total_rss_gb": 0.8332,
         "note": "prefill 後的峰值；prefill 會摸遍全部 expert，所以 peak 不代表 decode"
       },
       "rss_timeline_mib": [
         [
           0.0,
-          0.5
-        ],
-        [
-          0.205,
-          75.9
+          0.6
         ],
         [
-          0.41,
-          72.6
+          0.208,
+          75.7
         ],
         [
-          0.623,
-          135.2
+          0.414,
+          76.2
         ],
         [
-          0.831,
-          189.9
+          0.622,
+          141.0
         ],
         [
-          1.041,
-          247.8
+          0.834,
+          200.4
         ],
         [
-          1.251,
-          306.1
+          1.047,
+          261.3
         ],
         [
-          1.463,
-          364.3
+          1.261,
+          317.8
         ],
         [
-          1.676,
-          424.6
+          1.477,
+          380.5
         ],
         [
-          1.892,
-          482.6
+          1.694,
+          439.4
         ],
         [
-          2.117,
-          604.1
+          1.914,
+          509.0
         ],
         [
-          2.341,
-          688.7
+          2.144,
+          646.4
         ],
         [
-          2.564,
-          746.2
+          2.395,
+          739.6
         ],
         [
-          2.787,
-          828.9
+          2.645,
+          731.2
         ],
         [
-          3.025,
-          913.8
+          2.889,
+          716.3
         ],
         [
-          3.254,
-          994.6
+          3.136,
+          703.3
         ],
         [
-          3.481,
-          1080.7
+          3.406,
+          714.2
         ],
         [
-          3.711,
-          1166.4
+          3.662,
+          709.8
         ],
         [
-          3.941,
-          1236.5
+          3.907,
+          707.7
         ],
         [
-          4.18,
-          1228.6
+          4.133,
+          706.1
         ],
         [
-          4.416,
-          1241.3
+          4.357,
+          714.0
         ],
         [
-          4.65,
-          1204.6
+          4.582,
+          717.4
         ],
         [
-          4.888,
-          1212.4
+          4.806,
+          714.7
         ],
         [
-          5.128,
-          1250.3
+          5.031,
+          711.4
         ],
         [
-          5.368,
-          1211.6
+          5.258,
+          709.2
         ],
         [
-          5.605,
-          1218.1
+          5.486,
+          711.9
         ],
         [
-          5.837,
-          1200.6
+          5.709,
+          691.6
         ],
         [
-          6.07,
-          1248.6
+          5.933,
+          732.7
         ],
         [
-          6.304,
-          1271.8
+          6.156,
+          756.3
         ],
         [
-          6.55,
-          1315.0
+          6.389,
+          812.4
         ],
         [
-          6.794,
-          1322.1
+          6.627,
+          834.9
         ],
         [
-          7.043,
-          1329.8
+          6.865,
+          846.6
         ],
         [
-          7.304,
-          1326.6
+          7.1,
+          847.0
         ],
         [
-          7.553,
-          475.5
+          7.356,
+          852.7
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 8.1 t/s | Generation: 24.2 t/s ]",
+      "timing_line": "[ Prompt: 7.9 t/s | Generation: 23.1 t/s ]",
       "output": "Hello! 😊 I'm DeepSeek-R1, your friendly AI assistant. I'm here to help with all kinds of questions, learning, and creative tasks",
       "pager": {
         "routing_calls": 8448,
         "expert_touches": 64544,
-        "hits": 37015,
-        "misses": 2273,
-        "evictions": 2894,
-        "evicted_bytes": 6197382144,
-        "sweeps": 864,
-        "resident_bytes": 863972352,
-        "budget_bytes": 864026624,
-        "hit_rate": 0.5735,
-        "proc_self_read_bytes": 2429562880,
+        "hits": 35587,
+        "misses": 3701,
+        "evictions": 4572,
+        "evicted_bytes": 9822062592,
+        "sweeps": 1046,
+        "resident_bytes": 327075840,
+        "budget_bytes": 327155712,
+        "hit_rate": 0.5514,
+        "proc_self_read_bytes": 2426560512,
         "proc_self_rchar": 11947223
       },
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true,
-      "rss_vs_baseline": -1.2633,
-      "decode_rss_vs_baseline": -1.2633,
+      "rss_vs_baseline": -1.731,
+      "decode_rss_vs_baseline": -1.731,
       "decode_rss_lower_than_baseline": true,
       "rss_lower_than_baseline": true,
-      "ssd_reads_mib": 2317.0,
+      "ssd_reads_mib": 2314.1,
       "cold_weights_really_read_from_ssd": true,
-      "pager_hit_rate": 0.5735,
-      "pager_evictions": 2894,
-      "pager_misses": 2273,
-      "pager_resident_mib": 823.9,
-      "pager_budget_mib": 824.0
+      "pager_hit_rate": 0.5514,
+      "pager_evictions": 4572,
+      "pager_misses": 3701,
+      "pager_resident_mib": 311.9,
+      "pager_budget_mib": 312.0
     },
-    "budget512mb": {
+    "budget256mb": {
       "env": {
         "ST_PAGER": "1",
-        "ST_RAM_BUDGET_MB": "512",
+        "ST_RAM_BUDGET_MB": "256",
         "ST_RESERVE_MB": "200"
       },
       "rc": 0,
-      "seconds": 7.79,
+      "seconds": 7.77,
       "peak": {
-        "total_rss_gb": 0.8316,
-        "anon_rss_gb": 0.103,
-        "file_rss_gb": 0.8276,
+        "total_rss_gb": 0.6013,
+        "anon_rss_gb": 0.1029,
+        "file_rss_gb": 0.6017,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 0.8316
+        "hwm_rss_gb": 0.6013
       },
       "decode": {
-        "total_rss_gb": 0.8316,
+        "total_rss_gb": 0.6013,
         "note": "prefill 後的峰值；prefill 會摸遍全部 expert，所以 peak 不代表 decode"
       },
       "rss_timeline_mib": [
         [
           0.0,
-          0.6
+          0.4
         ],
         [
-          0.205,
-          75.8
+          0.206,
+          76.1
         ],
         [
-          0.409,
-          76.5
+          0.412,
+          71.5
         ],
         [
-          0.627,
-          141.0
+          0.619,
+          133.1
         ],
         [
-          0.834,
-          202.7
+          0.828,
+          188.8
         ],
         [
-          1.043,
-          261.2
+          1.039,
+          211.1
         ],
         [
-          1.252,
-          319.1
+          1.25,
+          224.4
         ],
         [
-          1.464,
-          372.0
+          1.46,
+          246.7
         ],
         [
-          1.679,
-          425.2
+          1.671,
+          270.4
         ],
         [
-          1.894,
-          479.1
+          1.885,
+          285.9
         ],
         [
-          2.113,
-          602.5
+          2.102,
+          363.2
         ],
         [
-          2.334,
-          686.2
+          2.325,
+          483.0
         ],
         [
-          2.559,
-          721.0
+          2.549,
+          480.7
         ],
         [
-          2.781,
-          694.2
+          2.769,
+          456.0
         ],
         [
-          3.008,
-          729.6
+          2.986,
+          435.9
         ],
         [
-          3.231,
-          711.1
+          3.205,
+          481.1
         ],
         [
-          3.456,
-          740.1
+          3.423,
+          462.3
         ],
         [
-          3.69,
-          738.7
+          3.64,
+          444.1
         ],
         [
-          3.92,
-          721.3
+          3.858,
+          477.7
         ],
         [
-          4.143,
-          712.6
+          4.077,
+          464.8
         ],
         [
-          4.367,
-          711.4
+          4.3,
+          461.9
         ],
         [
-          4.595,
-          728.2
+          4.524,
+          470.7
         ],
         [
-          4.825,
-          728.1
+          4.743,
+          470.7
         ],
         [
-          5.049,
-          726.2
+          4.963,
+          465.0
         ],
         [
-          5.274,
-          724.0
+          5.185,
+          457.7
         ],
         [
-          5.5,
-          722.1
+          5.404,
+          456.4
         ],
         [
-          5.749,
-          729.2
+          5.623,
+          460.3
         ],
         [
-          5.981,
-          711.0
+          5.841,
+          444.7
         ],
         [
-          6.208,
-          748.3
+          6.06,
+          483.1
         ],
         [
-          6.436,
-          770.3
+          6.28,
+          511.2
         ],
         [
-          6.667,
-          817.9
+          6.51,
+          563.8
         ],
         [
-          6.898,
-          837.3
+          6.743,
+          593.6
         ],
         [
-          7.127,
-          843.5
+          6.987,
+          603.9
         ],
         [
-          7.356,
-          840.8
+          7.234,
+          608.2
         ],
         [
-          7.584,
-          844.6
+          7.479,
+          614.6
+        ],
+        [
+          7.733,
+          77.2
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 7.9 t/s | Generation: 23.3 t/s ]",
+      "timing_line": "[ Prompt: 8.1 t/s | Generation: 21.6 t/s ]",
       "output": "Hello! 😊 I'm DeepSeek-R1, your friendly AI assistant. I'm here to help with all kinds of questions, learning, and creative tasks",
       "pager": {
         "routing_calls": 8448,
         "expert_touches": 64544,
-        "hits": 35587,
-        "misses": 3701,
-        "evictions": 4572,
-        "evicted_bytes": 9822062592,
-        "sweeps": 1046,
-        "resident_bytes": 327075840,
-        "budget_bytes": 327155712,
-        "hit_rate": 0.5514,
-        "proc_self_read_bytes": 2427265024,
+        "hits": 34660,
+        "misses": 4628,
+        "evictions": 5625,
+        "evicted_bytes": 12070591488,
+        "sweeps": 1280,
+        "resident_bytes": 57701376,
+        "budget_bytes": 58720256,
+        "hit_rate": 0.537,
+        "proc_self_read_bytes": 2432499712,
         "proc_self_rchar": 11947223
       },
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true,
-      "rss_vs_baseline": -1.7328,
-      "decode_rss_vs_baseline": -1.7328,
+      "rss_vs_baseline": -1.9629,
+      "decode_rss_vs_baseline": -1.9629,
       "decode_rss_lower_than_baseline": true,
       "rss_lower_than_baseline": true,
-      "ssd_reads_mib": 2314.8,
+      "ssd_reads_mib": 2319.8,
       "cold_weights_really_read_from_ssd": true,
-      "pager_hit_rate": 0.5514,
-      "pager_evictions": 4572,
-      "pager_misses": 3701,
-      "pager_resident_mib": 311.9,
-      "pager_budget_mib": 312.0
+      "pager_hit_rate": 0.537,
+      "pager_evictions": 5625,
+      "pager_misses": 4628,
+      "pager_resident_mib": 55.0,
+      "pager_budget_mib": 56.0
     },
-    "budget256mb": {
+    "budget128mb": {
       "env": {
         "ST_PAGER": "1",
-        "ST_RAM_BUDGET_MB": "256",
+        "ST_RAM_BUDGET_MB": "128",
         "ST_RESERVE_MB": "200"
       },
       "rc": 0,
-      "seconds": 7.97,
+      "seconds": 7.69,
       "peak": {
-        "total_rss_gb": 0.5875,
+        "total_rss_gb": 0.6643,
         "anon_rss_gb": 0.1029,
-        "file_rss_gb": 0.5864,
+        "file_rss_gb": 0.6689,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 0.5875
+        "hwm_rss_gb": 0.6643
       },
       "decode": {
-        "total_rss_gb": 0.5875,
+        "total_rss_gb": 0.6643,
         "note": "prefill 後的峰值；prefill 會摸遍全部 expert，所以 peak 不代表 decode"
       },
       "rss_timeline_mib": [
@@ -605,176 +597,164 @@
         ],
         [
           0.206,
-          76.1
+          76.2
         ],
         [
-          0.411,
-          72.7
+          0.412,
+          71.5
         ],
         [
           0.618,
-          140.0
-        ],
-        [
-          0.826,
-          192.9
-        ],
-        [
-          1.034,
-          214.1
+          134.8
         ],
         [
-          1.244,
-          233.8
+          0.827,
+          189.1
         ],
         [
-          1.456,
-          250.4
+          1.038,
+          248.9
         ],
         [
-          1.667,
-          271.8
-        ],
-        [
-          1.881,
-          284.4
+          1.251,
+          306.3
         ],
         [
-          2.094,
-          371.8
+          1.464,
+          319.8
         ],
         [
-          2.312,
-          482.8
+          1.681,
+          346.8
         ],
         [
-          2.531,
-          471.7
+          1.903,
+          361.4
         ],
         [
-          2.753,
-          439.1
+          2.128,
+          489.9
         ],
         [
-          2.976,
-          468.5
+          2.364,
+          555.3
         ],
         [
-          3.2,
-          444.3
+          2.601,
+          557.8
         ],
         [
-          3.425,
-          486.4
+          2.84,
+          555.9
         ],
         [
-          3.649,
-          478.2
+          3.077,
+          539.4
         ],
         [
-          3.883,
-          459.5
+          3.316,
+          528.6
         ],
         [
-          4.105,
-          448.2
+          3.56,
+          535.8
         ],
         [
-          4.33,
-          440.3
+          3.801,
+          538.3
         ],
         [
-          4.553,
-          448.0
+          4.038,
+          530.6
         ],
         [
-          4.784,
-          452.0
+          4.276,
+          542.2
         ],
         [
-          5.024,
-          452.2
+          4.515,
+          525.9
         ],
         [
-          5.244,
-          438.6
+          4.755,
+          529.8
         ],
         [
-          5.465,
-          436.2
+          4.997,
+          522.7
         ],
         [
-          5.686,
-          438.4
+          5.239,
+          538.4
         ],
         [
-          5.909,
-          484.5
+          5.482,
+          543.8
         ],
         [
-          6.139,
-          460.0
+          5.724,
+          546.0
         ],
         [
-          6.365,
-          489.1
+          5.967,
+          536.0
         ],
         [
-          6.596,
-          558.0
+          6.211,
+          581.3
         ],
         [
-          6.834,
-          576.9
+          6.46,
+          637.1
         ],
         [
-          7.071,
-          592.4
+          6.713,
+          660.9
         ],
         [
-          7.325,
-          594.1
+          6.968,
+          669.3
         ],
         [
-          7.561,
-          599.6
+          7.222,
+          679.7
         ],
         [
-          7.798,
-          594.5
+          7.473,
+          675.8
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 7.9 t/s | Generation: 20.3 t/s ]",
+      "timing_line": "[ Prompt: 8.3 t/s | Generation: 21.8 t/s ]",
       "output": "Hello! 😊 I'm DeepSeek-R1, your friendly AI assistant. I'm here to help with all kinds of questions, learning, and creative tasks",
       "pager": {
         "routing_calls": 8448,
         "expert_touches": 64544,
-        "hits": 34660,
-        "misses": 4628,
-        "evictions": 5625,
-        "evicted_bytes": 12070591488,
-        "sweeps": 1280,
-        "resident_bytes": 57701376,
-        "budget_bytes": 58720256,
-        "hit_rate": 0.537,
-        "proc_self_read_bytes": 2428882944,
+        "hits": 34952,
+        "misses": 4336,
+        "evictions": 5298,
+        "evicted_bytes": 11377677312,
+        "sweeps": 1054,
+        "resident_bytes": 132240384,
+        "budget_bytes": 134217728,
+        "hit_rate": 0.5415,
+        "proc_self_read_bytes": 2429865984,
         "proc_self_rchar": 11947223
       },
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true,
-      "rss_vs_baseline": -1.9769,
-      "decode_rss_vs_baseline": -1.9769,
+      "rss_vs_baseline": -1.8999,
+      "decode_rss_vs_baseline": -1.8999,
       "decode_rss_lower_than_baseline": true,
       "rss_lower_than_baseline": true,
-      "ssd_reads_mib": 2316.4,
+      "ssd_reads_mib": 2317.3,
       "cold_weights_really_read_from_ssd": true,
-      "pager_hit_rate": 0.537,
-      "pager_evictions": 5625,
-      "pager_misses": 4628,
-      "pager_resident_mib": 55.0,
-      "pager_budget_mib": 56.0
+      "pager_hit_rate": 0.5415,
+      "pager_evictions": 5298,
+      "pager_misses": 4336,
+      "pager_resident_mib": 126.1,
+      "pager_budget_mib": 128.0
     }
   },
   "all_passed": true