v4: 修 prefetch 反轉 bug — peak RSS 2.56→0.59 GiB，輸出逐字相同，all_passed=true

patches/0001-st-expert-pager.patch

@@ -1,23 +1,12 @@
-From 90deb52647021d7b6d0ff0d97a242f959fc6dce1 Mon Sep 17 00:00:00 2001
-From: c <a@b>
-Date: Wed, 7 Oct 2026 16:15:13 +0200
-Subject: [PATCH] st pager: disable MAP_POPULATE when paging, fix /proc/self/io
- parser
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: HelloSun <agent@huggingface.co>
+Date: Thu, 7 Oct 2026 17:00:00 +0200
+Subject: [PATCH] st: SmallThinker expert 級 SSD 分頁
 
----
- .st-patched         |   0
- .st-unpatched       |   0
- src/CMakeLists.txt  |   1 +
- src/llama-graph.cpp |  10 +
- src/llama-model.cpp |  41 +++-
- src/st_pager.cpp    | 535 ++++++++++++++++++++++++++++++++++++++++++++
- src/st_pager.h      |  80 +++++++
- 7 files changed, 666 insertions(+), 1 deletion(-)
- create mode 100644 .st-patched
- create mode 100644 .st-unpatched
- create mode 100644 src/st_pager.cpp
- create mode 100644 src/st_pager.h
+熱的 expert 權重留在 RAM，冷的用 madvise + posix_fadvise 丟回 SSD。
+計算路徑完全不動（仍是 llama.cpp 自己的 mul_mat_id），所以輸出與上游逐字相同。
 
+---
 diff --git a/.st-patched b/.st-patched
 new file mode 100644
 index 000000000..e69de29bb
@@ -64,7 +53,7 @@ index 1112ad885..e335a9d5f 100644
  
      if (weight_before_ffn) {
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index 1f3b80b08..a74118ae5 100644
+index 1f3b80b08..662491537 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
 @@ -1,5 +1,7 @@
@@ -84,22 +73,50 @@ index 1f3b80b08..a74118ae5 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 +1825,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -1821,7 +1825,24 @@ 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
 +    // 就已經超過任何分頁預算了，後面的 madvise 再怎麼丟都沒用（峰值已經發生）。
 +    // 這是量測方式／時機的坑，不是分頁機制本身有問題。
 +    // 註意 st::set_config_from_env() 必須在這之前呼叫（見 load_tensors 開頭）。
-+    ml.init_mappings(!st::prefetch_requested(), use_mlock ? &pimpl->mlock_mmaps : nullptr);
++    ml.init_mappings(!st::will_page(arch_name().c_str()), use_mlock ? &pimpl->mlock_mmaps : nullptr);
      pimpl->mappings.reserve(ml.mappings.size());
  
      // create the backend buffers
-@@ -1969,6 +1979,35 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -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) {
          }
      }
  
@@ -123,7 +140,20 @@ index 1f3b80b08..a74118ae5 100644
 +            by_name.emplace(nm, t);
 +        }
 +
-+        if (getenv("ST_VERBOSE") && !map_base) {
++        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());
 +        }
@@ -137,10 +167,10 @@ index 1f3b80b08..a74118ae5 100644
  
 diff --git a/src/st_pager.cpp b/src/st_pager.cpp
 new file mode 100644
-index 000000000..30d339827
+index 000000000..9cdbc0068
 --- /dev/null
 +++ b/src/st_pager.cpp
-@@ -0,0 +1,535 @@
+@@ -0,0 +1,579 @@
 +// ST expert pager —— 熱 expert 留在 RAM，冷 expert 丟回 SSD
 +//
 +// 設計取捨（為什麼不照抄 sddqwen35a3b_v01 的 pread arena）：
@@ -262,7 +292,15 @@ index 000000000..30d339827
 +
 +bool enabled() { return g_enabled && !g_slots.empty(); }
 +
-+bool prefetch_requested() { return !g_enabled; }
++// 這個查詢會在 init_mappings() **之前**被呼叫（那時還沒載入權重，
++// register_model 當然還沒跑過），所以不能只看 enabled()。
++// 它只能回答「這個模型**預計**會不會被分頁」。
++bool will_page(const char * arch) {
++    if (!g_enabled) return false;
++    return arch && strcmp(arch, "smallthinker") == 0;
++}
++
++bool paging_active() { return enabled(); }
 +
 +// ------------------------------------------------------------------ 統計
 +
@@ -485,6 +523,37 @@ index 000000000..30d339827
 +
 +// ------------------------------------------------------------------ 淘汰
 +
++// 診斷用：行程自己的 RSS（總量 / 匿名 / 檔案對映）。
++// 這是「分頁到底有沒生效」唯一可信的數字 —— 外部量測取樣太粗，
++// 而且 prefill 會讓 peak 失真。
++struct self_rss { size_t total = 0, anon = 0, file = 0; };
++
++static self_rss read_self_rss() {
++    self_rss r;
++    FILE * f = fopen("/proc/self/status", "r");
++    if (f) {
++        char line[256];
++        while (fgets(line, sizeof line, f)) {
++            unsigned long long v;
++            if (sscanf(line, "VmRSS: %llu", &v) == 1) r.total = v * 1024;
++        }
++        fclose(f);
++    }
++    f = fopen("/proc/self/smaps_rollup", "r");
++    if (f) {
++        char line[256], k[64];
++        unsigned long long v;
++        while (fgets(line, sizeof line, f)) {
++            if (sscanf(line, "%63[^:]: %llu", k, &v) != 2) continue;
++            if (!strcmp(k, "Anonymous"))                             r.anon  = v * 1024;
++            else if (!strcmp(k, "Private_Clean") || !strcmp(k, "Private_Dirty")) r.file += v * 1024;
++        }
++        fclose(f);
++    }
++    return r;
++}
++
++
 +// ST_EVICT_MODE 診斷用（實測發現只做其中一步會導致輸出損毀，見 docs）：
 +//   both（預設）| madvise | fadvise | none
 +static int evict_mode() {
@@ -572,12 +641,17 @@ index 000000000..30d339827
 +    }
 +
 +    if (g_verbose && !to_evict.empty()) {
++        const self_rss rss = read_self_rss();
 +        fprintf(stderr,
-+            "[st] sweep #%llu：丟 %zu 個 expert（%.1f MiB）→ resident %.1f MiB / budget %.1f MiB\n",
++            "[st] sweep #%llu：丟 %zu 個 expert（%.1f MiB）→ 帳 resident %.1f MiB / budget %.1f MiB"
++            " ｜ 實際 RSS total %.1f MiB（anon %.1f + file %.1f）\n",
 +            (unsigned long long) g_st.sweeps, to_evict.size(),
 +            freed / (1024.0 * 1024.0),
 +            (resident - freed) / (1024.0 * 1024.0),
-+            g_budget / (1024.0 * 1024.0));
++            g_budget / (1024.0 * 1024.0),
++            rss.total / (1024.0 * 1024.0),
++            rss.anon  / (1024.0 * 1024.0),
++            rss.file  / (1024.0 * 1024.0));
 +    }
 +}
 +
@@ -679,10 +753,10 @@ index 000000000..30d339827
 
 diff --git a/src/st_pager.h b/src/st_pager.h
 new file mode 100644
-index 000000000..43852a968
+index 000000000..336e37de3
 --- /dev/null
 +++ b/src/st_pager.h
-@@ -0,0 +1,80 @@
+@@ -0,0 +1,84 @@
 +// ST expert pager —— 熱 expert 留在 RAM，冷 expert 丟回 SSD
 +//
 +// 對外介面。llama.cpp 端（llama-model.cpp / llama-graph.cpp）只include 這個檔案。
@@ -755,7 +829,11 @@ index 000000000..43852a968
 +                               int layer, int n_expert, int n_expert_used);
 +
 +bool   enabled();
-+bool   prefetch_requested();   // 分頁開著時必須關掉 MAP_POPULATE
++// 分頁有註冊成功 → 必須關掉 MAP_POPULATE（否則載入時就會 fault 整個檔案）
++bool   paging_active();
++// 只看「有沒有打算分頁」，不看權重有沒有載入。init_mappings() 在載入之前，
++// 那時 register_model 還沒跑，所以只能靠 arch 判斷。
++bool   will_page(const char * arch);
 +stats  get_stats();
 +void   dump_stats(const char * path);
 +

validate/ab-test.json

@@ -1,5 +1,5 @@
 {
-  "when": "2026-10-07T16:14:12+0200",
+  "when": "2026-10-07T16:26:13+0200",
   "model": "/root/work/models/SmallThinker-4B-A0.6B-Instruct.Q4_K.gguf",
   "model_size_mib": 2508,
   "prompt": "Hello, who are you?",
@@ -12,342 +12,543 @@
         "ST_PAGER": "0"
       },
       "rc": 0,
-      "seconds": 7.21,
+      "seconds": 9.03,
       "peak": {
-        "total_rss_gb": 2.3009,
-        "anon_rss_gb": 0.1027,
-        "file_rss_gb": 2.2911,
+        "total_rss_gb": 2.5644,
+        "anon_rss_gb": 0.1033,
+        "file_rss_gb": 2.5587,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 2.3009
+        "hwm_rss_gb": 2.5644
       },
       "decode": {
-        "total_rss_gb": 2.3009,
+        "total_rss_gb": 2.5644,
         "note": "prefill 後的峰值；prefill 會摸遍全部 expert，所以 peak 不代表 decode"
       },
       "rss_timeline_mib": [
         [
-          0.001,
-          0.9
+          0.0,
+          0.6
         ],
         [
-          0.208,
-          75.9
+          0.207,
+          72.8
         ],
         [
-          0.414,
-          76.5
+          0.413,
+          71.3
+        ],
+        [
+          0.618,
+          117.7
+        ],
+        [
+          0.824,
+          187.4
+        ],
+        [
+          1.03,
+          255.4
+        ],
+        [
+          1.238,
+          326.7
         ],
         [
-          0.624,
-          142.8
+          1.451,
+          396.2
         ],
         [
-          0.836,
-          205.3
+          1.66,
+          467.0
         ],
         [
-          1.054,
-          266.2
+          1.87,
+          537.0
         ],
         [
-          1.275,
-          329.5
+          2.08,
+          608.5
         ],
         [
-          1.498,
-          395.0
+          2.291,
+          680.3
+        ],
+        [
+          2.504,
+          752.1
+        ],
+        [
+          2.723,
+          826.1
+        ],
+        [
+          2.938,
+          899.4
         ],
         [
-          1.715,
-          455.4
+          3.154,
+          971.1
         ],
         [
-          1.943,
-          559.2
+          3.37,
+          1044.4
         ],
         [
-          2.179,
-          689.2
+          3.586,
+          1118.2
         ],
         [
-          2.408,
-          755.8
+          3.803,
+          1191.2
         ],
         [
-          2.633,
-          843.2
+          4.022,
+          1264.2
         ],
         [
-          2.858,
-          924.7
+          4.241,
+          1338.3
         ],
         [
-          3.085,
-          1003.3
+          4.462,
+          1412.8
         ],
         [
-          3.313,
-          1092.9
+          4.695,
+          1489.0
         ],
         [
-          3.545,
-          1178.2
+          4.924,
+          1565.8
         ],
         [
-          3.787,
-          1266.8
+          5.154,
+          1642.8
         ],
         [
-          4.02,
-          1360.0
+          5.382,
+          1718.1
         ],
         [
-          4.253,
-          1451.9
+          5.608,
+          1795.6
         ],
         [
-          4.489,
-          1543.3
+          5.835,
+          1871.4
         ],
         [
-          4.727,
-          1631.2
+          6.063,
+          1947.9
         ],
         [
-          4.967,
-          1720.4
+          6.291,
+          2024.5
         ],
         [
-          5.207,
-          1814.3
+          6.522,
+          2100.2
         ],
         [
-          5.449,
-          1908.3
+          6.755,
+          2177.5
         ],
         [
-          5.695,
-          2001.3
+          6.987,
+          2255.8
         ],
         [
-          5.95,
-          2094.1
+          7.221,
+          2334.3
         ],
         [
-          6.203,
-          2171.2
+          7.459,
+          2411.6
         ],
         [
-          6.474,
-          2256.3
+          7.696,
+          2487.6
         ],
         [
-          6.759,
-          2324.0
+          7.934,
+          2566.1
         ],
         [
-          7.073,
-          1201.9
+          8.173,
+          2620.2
+        ],
+        [
+          8.414,
+          2626.0
+        ],
+        [
+          8.665,
+          2626.0
+        ],
+        [
+          8.918,
+          2626.0
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 8.2 t/s | Generation: 30.6 t/s ]",
+      "timing_line": "[ Prompt: 198.1 t/s | Generation: 55.6 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",
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true
     },
-    "budget512mb": {
+    "budget1024mb": {
       "env": {
         "ST_PAGER": "1",
-        "ST_RAM_BUDGET_MB": "512",
+        "ST_RAM_BUDGET_MB": "1024",
         "ST_RESERVE_MB": "200"
       },
       "rc": 0,
-      "seconds": 9.23,
+      "seconds": 7.6,
       "peak": {
-        "total_rss_gb": 2.5559,
-        "anon_rss_gb": 0.1035,
-        "file_rss_gb": 2.5503,
+        "total_rss_gb": 1.3011,
+        "anon_rss_gb": 0.103,
+        "file_rss_gb": 1.3028,
         "peak_swap_gb": 0.0,
-        "hwm_rss_gb": 2.5559
+        "hwm_rss_gb": 1.3011
       },
       "decode": {
-        "total_rss_gb": 2.5559,
+        "total_rss_gb": 1.3011,
         "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
+          0.5
         ],
         [
           0.205,
           75.9
         ],
         [
-          0.409,
-          71.3
+          0.41,
+          72.6
         ],
         [
-          0.614,
-          121.4
+          0.623,
+          135.2
         ],
         [
-          0.82,
-          190.4
+          0.831,
+          189.9
         ],
         [
-          1.026,
-          258.7
+          1.041,
+          247.8
         ],
         [
-          1.233,
-          326.7
+          1.251,
+          306.1
         ],
         [
-          1.44,
-          397.2
+          1.463,
+          364.3
         ],
         [
-          1.649,
-          468.3
+          1.676,
+          424.6
         ],
         [
-          1.858,
-          538.3
+          1.892,
+          482.6
         ],
         [
-          2.068,
-          608.8
+          2.117,
+          604.1
         ],
         [
-          2.278,
-          680.3
+          2.341,
+          688.7
+        ],
+        [
+          2.564,
+          746.2
+        ],
+        [
+          2.787,
+          828.9
+        ],
+        [
+          3.025,
+          913.8
+        ],
+        [
+          3.254,
+          994.6
+        ],
+        [
+          3.481,
+          1080.7
+        ],
+        [
+          3.711,
+          1166.4
+        ],
+        [
+          3.941,
+          1236.5
+        ],
+        [
+          4.18,
+          1228.6
+        ],
+        [
+          4.416,
+          1241.3
+        ],
+        [
+          4.65,
+          1204.6
+        ],
+        [
+          4.888,
+          1212.4
+        ],
+        [
+          5.128,
+          1250.3
+        ],
+        [
+          5.368,
+          1211.6
+        ],
+        [
+          5.605,
+          1218.1
+        ],
+        [
+          5.837,
+          1200.6
+        ],
+        [
+          6.07,
+          1248.6
+        ],
+        [
+          6.304,
+          1271.8
+        ],
+        [
+          6.55,
+          1315.0
+        ],
+        [
+          6.794,
+          1322.1
+        ],
+        [
+          7.043,
+          1329.8
+        ],
+        [
+          7.304,
+          1326.6
+        ],
+        [
+          7.553,
+          475.5
+        ]
+      ],
+      "io_delta": {},
+      "timing_line": "[ Prompt: 8.1 t/s | Generation: 24.2 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,
+        "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,
+      "decode_rss_lower_than_baseline": true,
+      "rss_lower_than_baseline": true,
+      "ssd_reads_mib": 2317.0,
+      "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
+    },
+    "budget512mb": {
+      "env": {
+        "ST_PAGER": "1",
+        "ST_RAM_BUDGET_MB": "512",
+        "ST_RESERVE_MB": "200"
+      },
+      "rc": 0,
+      "seconds": 7.79,
+      "peak": {
+        "total_rss_gb": 0.8316,
+        "anon_rss_gb": 0.103,
+        "file_rss_gb": 0.8276,
+        "peak_swap_gb": 0.0,
+        "hwm_rss_gb": 0.8316
+      },
+      "decode": {
+        "total_rss_gb": 0.8316,
+        "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
         ],
         [
-          2.49,
-          751.8
+          0.205,
+          75.8
+        ],
+        [
+          0.409,
+          76.5
+        ],
+        [
+          0.627,
+          141.0
         ],
         [
-          2.702,
-          824.6
+          0.834,
+          202.7
         ],
         [
-          2.916,
-          895.9
+          1.043,
+          261.2
         ],
         [
-          3.13,
-          968.6
+          1.252,
+          319.1
         ],
         [
-          3.346,
-          1041.7
+          1.464,
+          372.0
         ],
         [
-          3.563,
-          1115.7
+          1.679,
+          425.2
         ],
         [
-          3.78,
-          1187.2
+          1.894,
+          479.1
         ],
         [
-          3.998,
-          1262.2
+          2.113,
+          602.5
         ],
         [
-          4.216,
-          1335.2
+          2.334,
+          686.2
         ],
         [
-          4.436,
-          1409.3
+          2.559,
+          721.0
         ],
         [
-          4.659,
-          1484.6
+          2.781,
+          694.2
         ],
         [
-          4.881,
-          1558.8
+          3.008,
+          729.6
         ],
         [
-          5.103,
-          1633.1
+          3.231,
+          711.1
         ],
         [
-          5.326,
-          1706.4
+          3.456,
+          740.1
         ],
         [
-          5.551,
-          1780.9
+          3.69,
+          738.7
         ],
         [
-          5.777,
-          1856.4
+          3.92,
+          721.3
         ],
         [
-          6.003,
-          1931.9
+          4.143,
+          712.6
         ],
         [
-          6.239,
-          2009.0
+          4.367,
+          711.4
         ],
         [
-          6.469,
-          2085.5
+          4.595,
+          728.2
         ],
         [
-          6.7,
-          2161.5
+          4.825,
+          728.1
         ],
         [
-          6.933,
-          2239.5
+          5.049,
+          726.2
         ],
         [
-          7.182,
-          2317.5
+          5.274,
+          724.0
         ],
         [
-          7.418,
-          2400.6
+          5.5,
+          722.1
         ],
         [
-          7.657,
-          2480.1
+          5.749,
+          729.2
         ],
         [
-          7.895,
-          2558.6
+          5.981,
+          711.0
         ],
         [
-          8.128,
-          931.1
+          6.208,
+          748.3
         ],
         [
-          8.358,
-          935.8
+          6.436,
+          770.3
         ],
         [
-          8.584,
-          1119.4
+          6.667,
+          817.9
         ],
         [
-          8.825,
-          1193.9
+          6.898,
+          837.3
         ],
         [
-          9.057,
-          1193.3
+          7.127,
+          843.5
+        ],
+        [
+          7.356,
+          840.8
+        ],
+        [
+          7.584,
+          844.6
         ]
       ],
       "io_delta": {},
-      "timing_line": "[ Prompt: 128.4 t/s | Generation: 44.2 t/s ]",
+      "timing_line": "[ Prompt: 7.9 t/s | Generation: 23.3 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",
       "pager": {
         "routing_calls": 8448,
@@ -360,24 +561,221 @@
         "resident_bytes": 327075840,
         "budget_bytes": 327155712,
         "hit_rate": 0.5514,
-        "proc_self_read_bytes": 2630213632,
+        "proc_self_read_bytes": 2427265024,
         "proc_self_rchar": 11947223
       },
       "output_matches_baseline": true,
       "swap_zero": true,
       "rc_ok": true,
-      "rss_vs_baseline": 0.255,
-      "decode_rss_vs_baseline": 0.255,
-      "decode_rss_lower_than_baseline": false,
-      "rss_lower_than_baseline": false,
-      "ssd_reads_mib": 2508.4,
+      "rss_vs_baseline": -1.7328,
+      "decode_rss_vs_baseline": -1.7328,
+      "decode_rss_lower_than_baseline": true,
+      "rss_lower_than_baseline": true,
+      "ssd_reads_mib": 2314.8,
       "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
+    },
+    "budget256mb": {
+      "env": {
+        "ST_PAGER": "1",
+        "ST_RAM_BUDGET_MB": "256",
+        "ST_RESERVE_MB": "200"
+      },
+      "rc": 0,
+      "seconds": 7.97,
+      "peak": {
+        "total_rss_gb": 0.5875,
+        "anon_rss_gb": 0.1029,
+        "file_rss_gb": 0.5864,
+        "peak_swap_gb": 0.0,
+        "hwm_rss_gb": 0.5875
+      },
+      "decode": {
+        "total_rss_gb": 0.5875,
+        "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,
+          0.4
+        ],
+        [
+          0.206,
+          76.1
+        ],
+        [
+          0.411,
+          72.7
+        ],
+        [
+          0.618,
+          140.0
+        ],
+        [
+          0.826,
+          192.9
+        ],
+        [
+          1.034,
+          214.1
+        ],
+        [
+          1.244,
+          233.8
+        ],
+        [
+          1.456,
+          250.4
+        ],
+        [
+          1.667,
+          271.8
+        ],
+        [
+          1.881,
+          284.4
+        ],
+        [
+          2.094,
+          371.8
+        ],
+        [
+          2.312,
+          482.8
+        ],
+        [
+          2.531,
+          471.7
+        ],
+        [
+          2.753,
+          439.1
+        ],
+        [
+          2.976,
+          468.5
+        ],
+        [
+          3.2,
+          444.3
+        ],
+        [
+          3.425,
+          486.4
+        ],
+        [
+          3.649,
+          478.2
+        ],
+        [
+          3.883,
+          459.5
+        ],
+        [
+          4.105,
+          448.2
+        ],
+        [
+          4.33,
+          440.3
+        ],
+        [
+          4.553,
+          448.0
+        ],
+        [
+          4.784,
+          452.0
+        ],
+        [
+          5.024,
+          452.2
+        ],
+        [
+          5.244,
+          438.6
+        ],
+        [
+          5.465,
+          436.2
+        ],
+        [
+          5.686,
+          438.4
+        ],
+        [
+          5.909,
+          484.5
+        ],
+        [
+          6.139,
+          460.0
+        ],
+        [
+          6.365,
+          489.1
+        ],
+        [
+          6.596,
+          558.0
+        ],
+        [
+          6.834,
+          576.9
+        ],
+        [
+          7.071,
+          592.4
+        ],
+        [
+          7.325,
+          594.1
+        ],
+        [
+          7.561,
+          599.6
+        ],
+        [
+          7.798,
+          594.5
+        ]
+      ],
+      "io_delta": {},
+      "timing_line": "[ Prompt: 7.9 t/s | Generation: 20.3 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",
+      "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,
+        "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,
+      "decode_rss_lower_than_baseline": true,
+      "rss_lower_than_baseline": true,
+      "ssd_reads_mib": 2316.4,
+      "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
     }
   },
-  "all_passed": false
+  "all_passed": true
 }