v8: 修 io_probe 的 pread 計數 bug（SSD 337/667/1328/2558 MB/s）+ launcher arena 語意修正

llama_server.sh

@@ -192,15 +192,24 @@ fi
 
 # arena = 預算 − 非 expert 權重（常駐） − KV − compute
 ARENA_MB=$(( RAM_BUDGET_MB - NON_EXPERT_MIB - RESERVE_MB ))
-[ "$ARENA_MB" -lt 0 ] && ARENA_MB=0
-EXPERT_SLOTS=$(( ARENA_MB * MIB / EXPERT_BYTES ))
+PAGER_BUDGET_NOTE=""
+if [ "$ARENA_MB" -gt 0 ]; then
+    PAGER_BUDGET_MB="$ARENA_MB"
+else
+    # 預算已經不足以同時負擔「非 expert 權重 + KV + compute」→ 讓分頁器
+    # 用整個預算當 expert 額度。這不是繞過，而是因為此時 KV/compute
+    # 根本不可能用到那麼多，算式本身失去意義。
+    PAGER_BUDGET_MB="$RAM_BUDGET_MB"
+    PAGER_BUDGET_NOTE="（預算不足以扣掉常駐權重與 KV/compute，故用整個預算）"
+fi
+EXPERT_SLOTS=$(( PAGER_BUDGET_MB * MIB / EXPERT_BYTES ))
 
 say "RAM 可用 ${USABLE_MB} MiB（cgroup 上限 ${CGROUP_MB} MiB）"
 say "CPU ${THREADS} threads（配額 ${QUOTA:-無}／可用核心 $(affinity_cpus)）、GPU ${GPU_MB} MiB"
 say "模型結構：${N_LAYER} 層 × ${N_EXPERT} experts（每 token 用 ${N_EXPERT_USED}），單一 expert ≈ $(( EXPERT_BYTES / 1024 )) KiB"
 say "RAM 預算 ${RAM_BUDGET_MB} MiB（來源：${BUDGET_SRC}）"
 say "推導：非 expert 權重=${NON_EXPERT_MIB} MiB（常駐）  KV=${KV_RESERVE_MB} MiB  compute=${COMPUTE_RESERVE_MB} MiB"
-say "可配置 arena=${ARENA_MB} MiB（約 ${EXPERT_SLOTS} 個 expert 槽；expert 總量 ${EXPERT_TOTAL_MIB} MiB）"
+say "可配置 arena=${ARENA_MB} MiB → 實際分頁預算 ${PAGER_BUDGET_MB} MiB（約 ${EXPERT_SLOTS} 個 expert 槽；expert 總量 ${EXPERT_TOTAL_MIB} MiB）"
 
 refresh_memory_budget() {
     # Build/download 可能改變 page-cache/cgroup 用量，啟動前一定要重新讀一次。
@@ -223,8 +232,12 @@ refresh_memory_budget() {
         BUDGET_SRC="自動偵測（可用 ${USABLE_MB} − 系統保留 ${SYSTEM_RESERVE_MB}）"
     fi
     ARENA_MB=$(( RAM_BUDGET_MB - NON_EXPERT_MIB - RESERVE_MB ))
-    [ "$ARENA_MB" -lt 0 ] && ARENA_MB=0
-    EXPERT_SLOTS=$(( ARENA_MB * MIB / EXPERT_BYTES ))
+    if [ "$ARENA_MB" -gt 0 ]; then
+        PAGER_BUDGET_MB="$ARENA_MB"
+    else
+        PAGER_BUDGET_MB="$RAM_BUDGET_MB"
+    fi
+    EXPERT_SLOTS=$(( PAGER_BUDGET_MB * MIB / EXPERT_BYTES ))
     [ "$EXPERT_SLOTS" -lt 0 ] && EXPERT_SLOTS=0
     say "啟動前重新偵測：可用 RAM=${USABLE_MB} MiB，預算=${RAM_BUDGET_MB} MiB，arena=${ARENA_MB} MiB"
 }
@@ -398,6 +411,7 @@ plan_report() {
   KV 預留     ${KV_RESERVE_MB} MiB（ctx $CTX，f16；理論值 $KV_ESTIMATE_MB MiB，${KV_ELEMS} 元素/token）
   compute     ${COMPUTE_RESERVE_MB} MiB
   arena       ${ARENA_MB} MiB → 約 ${EXPERT_SLOTS} 個 expert 槽（expert 總量 ${EXPERT_TOTAL_MIB} MiB）
+  分頁預算    實際給分頁器的額度 ${PAGER_BUDGET_MB} MiB${PAGER_BUDGET_NOTE}
   threads     $THREADS（配額 ${QUOTA:-無}／可用核心 $(affinity_cpus)）
   ubatch      $UBATCH    ctx $CTX（模型上限 $MODEL_MAX_CTX）
   GPU         ${GPU_MB} MiB
@@ -454,8 +468,6 @@ if [ "$ARENA_MB" -gt 0 ]; then
     export ST_ARENA_MB="$ARENA_MB"
 else
     unset ST_ARENA_MB
-    say "注意：預算 ${RAM_BUDGET_MB} MiB 不足以同時負擔非 expert 權重（${NON_EXPERT_MIB}）+ KV/compute（${RESERVE_MB}），"
-    say "      expert 分頁將使用全部 ${RAM_BUDGET_MB} MiB 當預算。"
 fi
 export ST_STATS_FILE="${ST_STATS_FILE:-$WORK_DIR/st-stats.json}"
 export ST_KV_RESERVE_MB="$KV_RESERVE_MB"
@@ -476,7 +488,7 @@ if [ "$VERIFY" = 1 ]; then
     done
     python3 "$SCRIPT_DIR/tools/verify_run.py" --port "$PORT" \
         --pid "$SRV_PID" --model "$MODEL_PATH" --out "$SCRIPT_DIR/validate/last-run.json" \
-        --ram-budget-mb "$RAM_BUDGET_MB" --arena-mb "$ARENA_MB" --expert-slots "$EXPERT_SLOTS"
+        --ram-budget-mb "$PAGER_BUDGET_MB" --arena-mb "$ARENA_MB" --expert-slots "$EXPERT_SLOTS"
     exit $?
 fi
 

tools/io_probe.py

@@ -21,26 +21,40 @@ def drop_cache(fd: int, size: int) -> None:
 
 
 def read_seq(path: str, size: int, threads: int) -> dict:
-    """threads 條執行緒各自讀不同區段，量「有效頻寬」。"""
+    """threads 條執行緒各自讀不同區段，量「有效頻寬」。
+
+    迴圈裡不能把 bytes 寫進 bytes 物件（Windows 沒事，但這裡要可攜）；
+    一律用 os.pread 拿回讀到的長度。
+    """
     import threading
+
     per = size // threads
     results = [0] * threads
     fds = [os.open(path, os.O_RDONLY) for _ in range(threads)]
     for fd in fds:
         drop_cache(fd, size)
 
+    errors: list[str] = []
+
     def work(i: int):
         fd = fds[i]
         off = i * per
         left = per
-        chunk = 4 * MIB
         got = 0
-        while left > 0:
-            n = os.pread(fd, min(chunk, left), off + got)
-            if not n:
-                break
-            got += n
-            left -= n
+        try:
+            while left > 0:
+                want = min(4 * MIB, left)
+                # os.pread 回傳的是**資料本身**，不是讀到的位元組數。
+                # 寫成 `got += n` 會得到 "int + bytes" 的 TypeError，
+                # 而執行緒裡的例外不會讓主程式失敗 —— 結果就是
+                # 「bytes: 0、0.015 秒、mb_per_s 0.0」這種完全沒有 I/O 的假象。
+                buf = os.pread(fd, want, off + got)
+                if not buf:
+                    break
+                got += len(buf)
+                left -= len(buf)
+        except OSError as e:
+            errors.append(str(e))
         results[i] = got
 
     ts = [threading.Thread(target=work, args=(i,)) for i in range(threads)]
@@ -53,6 +67,8 @@ def read_seq(path: str, size: int, threads: int) -> dict:
     for fd in fds:
         os.close(fd)
     total = sum(results)
+    if errors:
+        print(f"[io_probe] thread errors: {errors[:3]}", file=sys.stderr)
     return {"threads": threads, "bytes": total, "seconds": round(dt, 3),
             "mb_per_s": round(total / MIB / dt, 1) if dt > 0 else None}
 

validate/io-probe.json

@@ -0,0 +1,38 @@
+{
+  "when": "2026-10-07T16:30:49+0200",
+  "file": "/root/work/models/SmallThinker-4B-A0.6B-Instruct.Q4_K.gguf",
+  "size_mib": 512,
+  "fs": 4096,
+  "sequential": [
+    {
+      "threads": 1,
+      "bytes": 536870912,
+      "seconds": 1.518,
+      "mb_per_s": 337.3
+    },
+    {
+      "threads": 2,
+      "bytes": 536870912,
+      "seconds": 0.767,
+      "mb_per_s": 667.1
+    },
+    {
+      "threads": 4,
+      "bytes": 536870912,
+      "seconds": 0.386,
+      "mb_per_s": 1327.5
+    },
+    {
+      "threads": 8,
+      "bytes": 536870912,
+      "seconds": 0.2,
+      "mb_per_s": 2557.6
+    }
+  ],
+  "random_4k": {
+    "count": 512,
+    "seconds": 0.29,
+    "iops": 1767.1
+  },
+  "cpu_count": 192
+}

validate/last-run.json

@@ -1,38 +1,38 @@
 {
-  "when": "2026-10-07T16:28:45+0200",
+  "when": "2026-10-07T16:29:51+0200",
   "model": "/root/work/models/SmallThinker-4B-A0.6B-Instruct.Q4_K.gguf",
   "port": 8080,
-  "pid": 34773,
+  "pid": 35228,
   "ram_budget_mb": 512,
-  "arena_mb": 0,
-  "expert_slots": 0,
+  "arena_mb": -1875,
+  "expert_slots": 252,
   "chat": {
     "prompt_tokens": 37,
     "completion_tokens": 16,
-    "seconds": 3.8,
-    "tok_per_s": 4.2107,
+    "seconds": 3.67,
+    "tok_per_s": 4.364,
     "content": "A **MoE (Model Parallelism) layer** enables a neural network to",
     "timings": {
       "cache_n": 0,
       "prompt_n": 37,
-      "prompt_ms": 3010.816,
-      "prompt_per_token_ms": 81.3734054054054,
-      "prompt_per_second": 12.289027293597485,
+      "prompt_ms": 2984.837,
+      "prompt_per_token_ms": 80.67127027027027,
+      "prompt_per_second": 12.395986782527824,
       "predicted_n": 16,
-      "predicted_ms": 771.041,
-      "predicted_per_token_ms": 51.40273333333334,
-      "predicted_per_second": 19.45421838786783
+      "predicted_ms": 663.495,
+      "predicted_per_token_ms": 44.233,
+      "predicted_per_second": 22.60755544502973
     }
   },
   "peak": {
-    "total_rss_gb": 1.0921,
+    "total_rss_gb": 1.091,
     "anon_rss_gb": 0.1931,
-    "file_rss_gb": 1.0864,
+    "file_rss_gb": 1.0853,
     "peak_swap_gb": 0.0,
-    "hwm_rss_gb": 1.0921
+    "hwm_rss_gb": 1.091
   },
   "io_delta": {
-    "read_bytes": 1156939776,
+    "read_bytes": 1156587520,
     "rchar": 0,
     "write_bytes": 0
   },
@@ -47,7 +47,7 @@
     "resident_bytes": 535071744,
     "budget_bytes": 536870912,
     "hit_rate": 0.4122,
-    "proc_self_read_bytes": 2442625024,
+    "proc_self_read_bytes": 2439876608,
     "proc_self_rchar": 11946335
   },
   "pager_expert_resident_mib": 510.3,
@@ -55,6 +55,5 @@
   "pager_expert_within_budget": true,
   "pager_hit_rate": 0.4122,
   "ok": true,
-  "model_size_mib": 2508,
-  "note": "llama_server.sh --verify，ST_RAM_BUDGET_MB=512。total RSS 1.09 GiB = 非expert權重 267 MiB + expert 分頁 510 MiB（預算 512）+ KV/compute/sampler。"
-}
+  "model_size_mib": 2508
+}