v1: 驗證通過 — llama-server 起來、chat 正常、peak RSS 4.30 GiB / swap 0

llama_server.sh

@@ -243,12 +243,19 @@ ensure_patches() {
         say "使用本地的 patch：$(cd "$PATCH_DIR" && ls *.patch | tr '\n' ' ')"
         return
     fi
+    # 遠端還沒有 patch（第一版就是這樣）是正常狀況，不是錯誤：
+    # 沒有 patch 時走上游行為，權重一樣是 SSD 上的 file-backed mmap。
     step "從 $HF_REPO 取得 patch"
-    local p
+    local p got=0
     for p in 0001-st-expert-pager.patch; do
-        hf_get "$HF_REPO" "patches/$p" "$PATCH_DIR/$p" && say "  patches/$p"
+        if curl -fsSL -H "Authorization: Bearer ${HF_TOKEN:-}" -o "$PATCH_DIR/$p.part" \
+             "https://huggingface.co/$HF_REPO/resolve/main/patches/$p" 2>/dev/null; then
+            mv -f "$PATCH_DIR/$p.part" "$PATCH_DIR/$p"; say "  patches/$p"; got=1
+        else
+            rm -f "$PATCH_DIR/$p.part"; say "  （遠端沒有 $p）"
+        fi
     done
-    [ -n "$(ls -A "$PATCH_DIR"/*.patch 2>/dev/null)" ] || say "  （這個 repo 目前沒有 patch，用上游行為）"
+    [ "$got" = 1 ] || say "  用上游行為：權重走 SSD file-backed mmap，沒有 expert 級分頁"
 }
 
 # ================================================== 3. 取得 llama.cpp 原始碼

validate/last-run.json

@@ -0,0 +1,41 @@
+{
+  "when": "2026-10-07T15:45:38+0200",
+  "model": "/root/work/models/SmallThinker-4B-A0.6B-Instruct.Q4_K.gguf",
+  "port": 8080,
+  "pid": 5409,
+  "ram_budget_mb": 92627,
+  "arena_mb": 90240,
+  "expert_slots": 44562,
+  "chat": {
+    "prompt_tokens": 37,
+    "completion_tokens": 16,
+    "seconds": 0.51,
+    "tok_per_s": 31.6285,
+    "content": "A **MoE (Model Parallelism) layer** enables a neural network to",
+    "timings": {
+      "cache_n": 0,
+      "prompt_n": 37,
+      "prompt_ms": 216.341,
+      "prompt_per_token_ms": 5.8470540540540545,
+      "prompt_per_second": 171.02629644866207,
+      "predicted_n": 16,
+      "predicted_ms": 255.155,
+      "predicted_per_token_ms": 17.01033333333333,
+      "predicted_per_second": 58.78779565362231
+    }
+  },
+  "peak": {
+    "total_rss_gb": 4.3048,
+    "anon_rss_gb": 1.8637,
+    "file_rss_gb": 4.2991,
+    "peak_swap_gb": 0.0,
+    "hwm_rss_gb": 4.3048
+  },
+  "io_delta": {
+    "read_bytes": 0,
+    "rchar": 6,
+    "write_bytes": 0
+  },
+  "model_size_mib": 2508,
+  "ok": true
+}