onw 1.2.2: API default sampling setting (for apps that send none), chat starts from custom values

CHANGELOG.md

@@ -5,6 +5,12 @@
 日付は日本時間です。各版は Hugging Face の `ryugyosoft/onw` にタグ（`v0.4.2` 以降）付きで残っています。
 onw ウィンドウは、新しい版があるとこのファイルから変更点を表示します（`## v…` の見出しで区切っています）。
 
+## v1.2.2 (2026-10-04)
+- 設定の「サーバーと API」に「API の既定の生成設定」を追加。温度・Top P などを送らないアプリ（OpenCode など）のリクエストに、
+  「貪欲（温度 0。これまでどおり、既定）」「モデルの推奨値（チャット画面と同じ。Qwen 系は思考時に思考用の値）」「自分で指定」の
+  どれを使うかを選べます。アプリが送った値はそのまま使います。「自分で指定」にすると、チャット画面の初期値もその値になります
+  （「（設定の値）」と表示）。
+
 ## v1.2.1 (2026-10-04)
 - Gemma：設定の「64 トークン処理」を切り替えるたびに、NPU 向けのコンパイル（数分）をやり直していた問題を修正。スライディング
   ウィンドウ層のリングの長さが「読み込んだ中で最大のブロック」で決まっていたため、64 トークン処理のオン／オフで 1 トークン用・

CHANGELOG_en.md

@@ -4,6 +4,12 @@
 
 Dates are JST. Every version is on Hugging Face `ryugyosoft/onw`, tagged from `v0.4.2` on.
 
+## v1.2.2 (2026-10-04)
+- Settings, Server and API: "API default sampling" for requests that send no temperature, Top P etc. (OpenCode, ...):
+  greedy (temperature 0, as before, the default), the model's recommended values (as in the chat; the Qwen family
+  switches to its thinking values when thinking) or custom values. Values an app sends are always used. With custom
+  values the chat page starts from them too ("(the values from the settings)").
+
 ## v1.2.1 (2026-10-04)
 - Gemma: switching "64-token blocks" in the settings compiled the model for the NPU again (minutes) every time. The
   sliding-window layers' ring length followed the largest block loaded, so turning the 64-token blocks on or off changed

MANUAL.md

@@ -143,6 +143,7 @@
 | コンテキスト長 | 会話で使える最大の長さ。KV キャッシュは実際に使った分だけメモリを使うので、大きくしても短い会話は遅くなりません。変えても再コンパイルはありません |
 | LAN からの接続を許可する | 他の PC から API を使えるようにします（API キーの設定をおすすめします） |
 | API で思考の指定がないリクエストも考えてから答える | `enable_thinking` を送らないエージェントのアプリ向け。回答は遅くなります |
+| API の既定の生成設定 | 温度・Top P などを送らないアプリ（OpenCode など）のリクエストに使う値。「貪欲（温度 0）」（既定）、「モデルの推奨値」（チャット画面と同じ。Qwen 系は思考時に思考用の値）、「自分で指定」。アプリが送った値はそのまま使います。「自分で指定」にすると、チャット画面の初期値もその値になります |
 
 ### 速さとメモリ
 

MANUAL_en.md

@@ -152,6 +152,7 @@ applies them ("Discard" throws them away). For settings that need a server resta
 | Context length | The longest a conversation can get. The KV cache only uses memory for what is used, so a larger value does not slow short conversations; changing it needs no recompile |
 | Allow connections from the LAN | Lets other PCs use the API (setting an API key is recommended) |
 | Think before answering API requests that do not say | For agent apps that never send `enable_thinking`. Answers take longer |
+| API default sampling | The values for requests that send no temperature, Top P etc. (OpenCode, ...): greedy (temperature 0, the default), the model's recommended values (as in the chat; the Qwen family switches to its thinking values when thinking) or custom. Values an app sends are always used. With custom values the chat page starts from them too |
 
 ### Speed and memory
 

onw/__init__.py

@@ -2,4 +2,4 @@
 
 A small engine that runs LLMs entirely on the Intel NPU: static segment graphs + host-managed state.
 """
-__version__ = "1.2.1"
+__version__ = "1.2.2"

onw/app.py

@@ -84,7 +84,8 @@ STREAMABLE = {
 }
 DEFAULTS = {"model": "", "port": 8000, "lan": False, "api_key": "", "context": 32768, "device": "NPU",
             "autostart_server": True, "login_start": False, "models_dir": "", "model_folders": [], "pld": True, "big_blocks": "auto", "npuw": "auto", "expert_gb": "auto", "update_check": True, "mem_saver": False, "mem_saver_strict": False, "think_default": False,
-            "lang": "auto", "npuw_moe": False, "npuw_async": False, "cache_size": False}
+            "lang": "auto", "npuw_moe": False, "npuw_async": False, "cache_size": False,
+            "api_sampling": "greedy", "api_temperature": 0.7, "api_top_p": 0.95, "api_top_k": 20, "api_presence": 0.0}
 CONTEXTS = ["1024", "2048", "4096", "8192", "16384", "32768", "65536", "131072", "262144"]
 
 
@@ -651,6 +652,14 @@ def _start_server(cfg):
         env["ONW_EXPERT_GB"] = str(budget)
     if cfg.get("think_default"):                      # API requests that say nothing about thinking: think
         env["ONW_THINK_DEFAULT"] = "1"
+    # API requests that send no sampling values (agent apps like OpenCode): greedy, the model's recommendation or set
+    if cfg.get("api_sampling") == "recommended":
+        env["ONW_API_SAMPLING"] = "recommended"
+    elif cfg.get("api_sampling") == "custom":
+        env["ONW_API_SAMPLING"] = json.dumps({"temperature": float(cfg.get("api_temperature", 0.7)),
+                                              "top_p": float(cfg.get("api_top_p", 0.95)),
+                                              "top_k": int(cfg.get("api_top_k", 0)),
+                                              "presence_penalty": float(cfg.get("api_presence", 0.0))})
     if cfg.get("mem_saver") and saver_applies(model, cfg):   # memory-saving mode: one block size at a time
         env["ONW_LAZY16"] = "1"
         if cfg.get("mem_saver_strict"):                # strict: never both block sizes at once (answers wait)

onw/i18n.json

@@ -10,6 +10,7 @@
 "35B MoE を半分に刈り込み": "The 35B MoE pruned by half",
 "64 トークン処理": "64-token blocks",
 "API で思考の指定がないリクエストも考えてから答える": "Think before answering API requests that do not say",
+"API の既定の生成設定": "API default sampling",
 "API キー（任意）": "API key (optional)",
 "Hugging Face に接続できません: {0}": "Cannot reach Hugging Face: {0}",
 "IBM の 8B。思考とツール呼び出し・コーディング向け": "IBM's 8B, for thinking, tool calls and coding",
@@ -132,6 +133,7 @@
 "メモリ目安": "Memory (est.)",
 "メモリ節約モード": "Memory-saving mode",
 "モデル": "Models",
+"モデルの推奨値（チャット画面と同じ）": "The model's recommended values (as in the chat)",
 "モデルは読み込まれていません": "No model loaded",
 "モデルをロード": "Load model",
 "モデルを削除": "Delete model",
@@ -253,9 +255,11 @@
 "次の文章を、元の表現をなるべく使って2文に要約してください。": "Summarize the following text in two sentences, reusing its wording as much as possible.",
 "止まった理由: {0}": "Why it stopped: {0}",
 "温度": "Temperature",
+"温度などを送らないアプリ（OpenCode など）のリクエストに使います。アプリが送った値はそのまま使います。": "For requests from apps that send no temperature etc. (OpenCode, ...). Values an app sends are always used.",
 "準備中…": "Preparing…",
 "無効": "Off",
 "状態": "Status",
+"生成設定の値が不正です": "Invalid sampling values",
 "生成速度": "Decode speed",
 "画像": "Image",
 "画像も読めます": "Reads images too",
@@ -267,6 +271,7 @@
 "終了（サーバーも停止）": "Quit (stops the server too)",
 "結果": "Result",
 "繰り返し抑制": "Repetition penalty",
+"自分で指定": "Custom",
 "自分で置いたモデル・{0}": "Your own model · {0}",
 "要約や書き直しなど、元の文章をなぞる回答が速くなります。": "Speeds up answers that reuse the prompt's text, such as summaries and rewrites.",
 "言語": "Language",
@@ -274,6 +279,7 @@
 "設定は保存しました。サーバーを新しい設定で今すぐ再起動しますか？（キャンセルしても、次にモデルを読み込むときから反映されます）": "Settings saved. Restart the server with the new settings now? (Cancel keeps them for the next model load.)",
 "詳細": "Advanced",
 "読み込み中…": "Loading…",
+"貪欲（温度 0。毎回同じ答え）": "Greedy (temperature 0, the same answer every time)",
 "起動時と 1 日ごとに更新を確認する": "Check for updates at startup and daily",
 "起動時に前回ロードしたモデルを自動でロードする": "Load the last model automatically at startup",
 "追加…": "Add…",
@@ -298,6 +304,7 @@
 "（再利用 {0}）": " (reused {0})",
 "（無題）": "(untitled)",
 "（画像）": "(image)",
+"（設定の値）": "(the values from the settings)",
 "＋ 新しい会話": "+ New conversation"
 },
 "zh": {
@@ -311,6 +318,7 @@
 "35B MoE を半分に刈り込み": "35B MoE 裁剪一半",
 "64 トークン処理": "64 token 分块",
 "API で思考の指定がないリクエストも考えてから答える": "API 请求未指定思考时也先思考再回答",
+"API の既定の生成設定": "API 默认生成设置",
 "API キー（任意）": "API 密钥（可选）",
 "Hugging Face に接続できません: {0}": "无法连接 Hugging Face：{0}",
 "IBM の 8B。思考とツール呼び出し・コーディング向け": "IBM 的 8B，适合思考、工具调用和编程",
@@ -433,6 +441,7 @@
 "メモリ目安": "内存估计",
 "メモリ節約モード": "省内存模式",
 "モデル": "模型",
+"モデルの推奨値（チャット画面と同じ）": "模型的推荐值（与聊天页面相同）",
 "モデルは読み込まれていません": "未加载模型",
 "モデルをロード": "加载模型",
 "モデルを削除": "删除模型",
@@ -554,9 +563,11 @@
 "次の文章を、元の表現をなるべく使って2文に要約してください。": "请尽量使用原文的表达，把下面的文字概括成两句话。",
 "止まった理由: {0}": "停止原因：{0}",
 "温度": "温度",
+"温度などを送らないアプリ（OpenCode など）のリクエストに使います。アプリが送った値はそのまま使います。": "用于不发送温度等参数的应用（如 OpenCode）的请求。应用发送的值始终优先。",
 "準備中…": "准备中…",
 "無効": "关闭",
 "状態": "状态",
+"生成設定の値が不正です": "生成设置的值无效",
 "生成速度": "生成速度",
 "画像": "图像",
 "画像も読めます": "也能读图",
@@ -568,6 +579,7 @@
 "終了（サーバーも停止）": "退出（同时停止服务器）",
 "結果": "结果",
 "繰り返し抑制": "重复惩罚",
+"自分で指定": "自定义",
 "自分で置いたモデル・{0}": "自行放置的模型・{0}",
 "要約や書き直しなど、元の文章をなぞる回答が速くなります。": "让摘要、改写等沿用原文的回答更快。",
 "言語": "语言",
@@ -575,6 +587,7 @@
 "設定は保存しました。サーバーを新しい設定で今すぐ再起動しますか？（キャンセルしても、次にモデルを読み込むときから反映されます）": "设置已保存。现在用新设置重启服务器？（取消后，下次加载模型时生效）",
 "詳細": "高级",
 "読み込み中…": "加载中…",
+"貪欲（温度 0。毎回同じ答え）": "贪婪（温度 0，每次答案相同）",
 "起動時と 1 日ごとに更新を確認する": "启动时及每天检查更新",
 "起動時に前回ロードしたモデルを自動でロードする": "启动时自动加载上次的模型",
 "追加…": "添加…",
@@ -599,6 +612,7 @@
 "（再利用 {0}）": "（复用 {0}）",
 "（無題）": "（无标题）",
 "（画像）": "（图片）",
+"（設定の値）": "（设置中的值）",
 "＋ 新しい会話": "＋ 新对话"
 }
 }

onw/manager.py

@@ -373,11 +373,21 @@ class Manager:
                   "update_check", "lang", "npuw_moe", "npuw_async", "cache_size", "mem_saver", "mem_saver_strict", "think_default"):
             if k in new:
                 cfg[k] = new[k]
+        if new.get("api_sampling") in ("greedy", "recommended", "custom"):
+            cfg["api_sampling"] = new["api_sampling"]
+        try:
+            for k, cast, lo, hi in (("api_temperature", float, 0, 2), ("api_top_p", float, 0.01, 1), ("api_top_k", int, 0, 1000),
+                                    ("api_presence", float, 0, 2)):
+                if k in new:
+                    cfg[k] = min(max(cast(float(new[k])), lo), hi)
+        except (TypeError, ValueError):
+            return {"error": T("生成設定の値が不正です")}
         if ctx != cfg["context"]:
             cfg["context_chosen"] = True                       # (the old 1024 default moves up; a choice stays)
         cfg["port"], cfg["context"] = port, ctx
         server_keys = ("lan", "api_key", "device", "pld", "big_blocks", "npuw", "expert_gb", "port", "context",
-                       "npuw_moe", "npuw_async", "cache_size", "mem_saver", "mem_saver_strict", "think_default")
+                       "npuw_moe", "npuw_async", "cache_size", "mem_saver", "mem_saver_strict", "think_default",
+                       "api_sampling", "api_temperature", "api_top_p", "api_top_k", "api_presence")
         app.save_config(cfg)
         i18n.current(cfg)                                  # messages switch language right away
         try:

onw/server.py

@@ -40,6 +40,7 @@ info = {}
 lock = threading.Lock()     # one generation at a time on the NPU
 WEB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "web")
 API_KEY = None              # --api-key: required as a Bearer token on /v1
+API_SAMPLING = None         # sampling for requests that send none (ONW_API_SAMPLING: "recommended" or a JSON of values)
 
 
 def oai_error(status, message, type_="invalid_request_error", code=None):
@@ -293,10 +294,15 @@ def gen_args(body):
     elif isinstance(choice, dict) and tools:              # a named function: offer just that one
         name = (choice.get("function") or {}).get("name")
         tools = [t for t in tools if (t.get("function") or {}).get("name") == name] or tools
-    return dict(max_new_tokens=mx, temperature=_num(body, "temperature", 0.0), top_p=_num(body, "top_p", 1.0),
+    # what the request leaves out: greedy, or the settings' API defaults (API_SAMPLING; thinking has its own set)
+    dflt = dict(API_SAMPLING or {})
+    if think and isinstance(dflt.get("think"), dict):
+        dflt.update(dflt["think"])
+    return dict(max_new_tokens=mx, temperature=_num(body, "temperature", dflt.get("temperature", 0.0)),
+                top_p=_num(body, "top_p", dflt.get("top_p", 1.0)),
                 stop=[stop] if isinstance(stop, str) else list(stop), seed=body.get("seed"), think=think,
-                top_k=_num(body, "top_k", 0, int), repetition_penalty=_num(body, "repetition_penalty", 1.0),
-                presence_penalty=_num(body, "presence_penalty", 0.0),
+                top_k=_num(body, "top_k", dflt.get("top_k", 0), int), repetition_penalty=_num(body, "repetition_penalty", 1.0),
+                presence_penalty=_num(body, "presence_penalty", dflt.get("presence_penalty", 0.0)),
                 frequency_penalty=_num(body, "frequency_penalty", 0.0), tools=tools)
 
 
@@ -515,7 +521,7 @@ def recommended_sampling(arch, engine_dir=None):
 
 
 def main(argv=None):
-    global pipe, MODEL_ID, API_KEY
+    global pipe, MODEL_ID, API_KEY, API_SAMPLING
     ap = argparse.ArgumentParser(prog="onw serve", description="chat UI + OpenAI-compatible API on the Intel NPU")
     ap.add_argument("model", help="onw model folder, or a Hugging Face repo id (downloaded into ./models)")
     ap.add_argument("--host", default="127.0.0.1")
@@ -554,6 +560,18 @@ def main(argv=None):
     # greedy Gemma on the NPU can fall into repeating one word (fp16 on the NPU tips a near tie; the same happens with
     # any onw version): the pages suggest a repetition penalty, the API default stays 1.0 (OpenAI behaviour)
     info.update(arch=arch, rep_hint=1.2 if arch.startswith("gemma") else None, sampling=recommended_sampling(arch, a.engine))
+    want = os.environ.get("ONW_API_SAMPLING")
+    if want == "recommended":
+        API_SAMPLING = {k: v for k, v in (info["sampling"] or {}).items() if k != "source"} or None
+    elif want:
+        try:
+            API_SAMPLING = json.loads(want)
+        except ValueError:
+            API_SAMPLING = None
+    info.update(api_sampling=API_SAMPLING, api_sampling_mode="recommended" if want == "recommended" else
+                "custom" if API_SAMPLING else "greedy")
+    if API_SAMPLING:
+        print(f"[api] requests without sampling values use {API_SAMPLING}", flush=True)
     if a.open:
         import webbrowser
         threading.Timer(1.5, webbrowser.open, [f"http://{'localhost' if a.host in ('0.0.0.0', '127.0.0.1') else a.host}:{a.port}/chat"]).start()

onw/web/index.html

@@ -195,7 +195,10 @@ fetch("/api/info").then(r => r.json()).then(i => {
   if (i.always_thinks && $("maxtok").value === "512") $("maxtok").value = "2048";
 
   if (i.context) $("note").textContent = T("文脈は約{0}トークン", i.context);
-  if (i.sampling) {                         // the model's recommended sampling as the starting values
+  if (i.api_sampling_mode === "custom" && i.api_sampling) {   // the settings' own API values: the chat starts there too
+    SAMPLING = i.api_sampling; setSampling(SAMPLING);
+    $("samphint").textContent = T("（設定の値）"); $("samphint").style.display = "";
+  } else if (i.sampling) {                  // the model's recommended sampling as the starting values
     SAMPLING = i.sampling; setSampling(SAMPLING);
     $("samphint").textContent = i.sampling.source === "onw" ? T("（onw の推奨値）") : T("（モデルの推奨値）");
     $("samphint").style.display = "";

onw/web/manage.html

@@ -217,6 +217,13 @@
         <span class="desc">会話で使える最大の長さ。KV キャッシュは実際に使った分だけメモリを使います（1K トークンあたり 8〜80 MB）。変えても再コンパイルはありません。</span></div>
       <label class="swrow"><span class="switch"><input type="checkbox" id="f-lan"><i></i></span><span>LAN からの接続を許可する<span class="desc">他の PC から API を使えるようにします。API キーの設定をおすすめします。</span></span></label>
       <label class="swrow"><span class="switch"><input type="checkbox" id="f-think_default"><i></i></span><span>API で思考の指定がないリクエストも考えてから答える<span class="desc">enable_thinking を送らないエージェントのアプリ向け。回答は遅くなります。</span></span></label>
+      <div class="opt"><span class="k">API の既定の生成設定</span><select id="f-api_sampling" style="width:260px"><option value="greedy">貪欲（温度 0。毎回同じ答え）</option><option value="recommended">モデルの推奨値（チャット画面と同じ）</option><option value="custom">自分で指定</option></select>
+        <span class="desc">温度などを送らないアプリ（OpenCode など）のリクエストに使います。アプリが送った値はそのまま使います。</span></div>
+      <div class="opt" id="row-api_custom"><span class="k"></span>
+        <label class="small">温度 <input type="number" id="f-api_temperature" min="0" max="2" step="0.1" style="width:64px"></label>
+        <label class="small">Top P <input type="number" id="f-api_top_p" min="0.05" max="1" step="0.05" style="width:64px"></label>
+        <label class="small">Top K <input type="number" id="f-api_top_k" min="0" max="1000" step="1" style="width:64px"></label>
+        <label class="small">出現ペナルティ <input type="number" id="f-api_presence" min="0" max="2" step="0.1" style="width:64px"></label></div>
     </div>
 
     <div class="card set sub" data-sub="speed">
@@ -391,6 +398,9 @@
       $("f-device").value = c.device; $("f-big_blocks").value = c.big_blocks || "auto"; $("f-npuw").value = c.npuw || "auto"; $("f-expert_gb").value = String(c.expert_gb || "auto"); $("f-lang").value = c.lang || "auto";
       for (const k of ["pld", "autostart_server", "login_start", "lan", "update_check"]) $("f-" + k).checked = c[k] !== false && !!(c[k] ?? true);
       for (const k of ["npuw_moe", "npuw_async", "cache_size", "mem_saver", "mem_saver_strict", "think_default"]) $("f-" + k).checked = !!c[k];
+      $("f-api_sampling").value = c.api_sampling || "greedy";
+      for (const k of ["api_temperature", "api_top_p", "api_top_k", "api_presence"]) $("f-" + k).value = c[k] ?? "";
+      apiCustomState();
       strictState();
       const sm = s.saver_models || [];
       $("row-mem_saver").style.display = sm.length || c.mem_saver ? "" : "none";
@@ -398,6 +408,8 @@
       $("saver-models").textContent = sm.length ? T("この PC で効くモデル：{0}", sm.join(T("、"))) : T("この PC には効くモデルがありません");
     }
   }
+  function apiCustomState() { $("row-api_custom").style.display = $("f-api_sampling").value === "custom" ? "" : "none"; }
+  document.addEventListener("change", e => { if (e.target && e.target.id === "f-api_sampling") apiCustomState(); });
   function strictState() {                   // the strict option only under the memory-saving mode
     const on = $("f-mem_saver").checked;
     $("f-mem_saver_strict").disabled = !on;
@@ -448,7 +460,9 @@
   }
   async function saveSettings() {
     const body = {port: $("f-port").value, context: $("f-context").value, api_key: $("f-api_key").value.trim(), device: $("f-device").value,
-                  big_blocks: $("f-big_blocks").value, npuw: $("f-npuw").value, expert_gb: $("f-expert_gb").value, lang: $("f-lang").value};
+                  big_blocks: $("f-big_blocks").value, npuw: $("f-npuw").value, expert_gb: $("f-expert_gb").value, lang: $("f-lang").value,
+                  api_sampling: $("f-api_sampling").value, api_temperature: $("f-api_temperature").value, api_top_p: $("f-api_top_p").value,
+                  api_top_k: $("f-api_top_k").value, api_presence: $("f-api_presence").value};
     for (const k of ["pld", "autostart_server", "login_start", "lan", "update_check", "npuw_moe", "npuw_async", "cache_size", "mem_saver", "mem_saver_strict", "think_default"]) body[k] = $("f-" + k).checked;
     if (!body.mem_saver) body.mem_saver_strict = false;
     const oldLang = S.config.lang || "auto";

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "onw"
-version = "1.2.1"
+version = "1.2.2"
 description = "Ore no NPU ga konna ni ugoku wake nai - LLMs entirely on the Intel NPU"
 readme = "README.md"
 license = {text = "Apache-2.0"}