feat(cmf): per-server O(1) attention override at serve time

Each managed local model row gains an O(1) selector (auto / off / all / deepN),
exported to the child `cortiq serve` as CMF_O1. "off" forces exact attention,
curing incoherent output from a too-aggressive `--o1` conversion hint without
re-converting the file; "all"/"deepN" force-enable it. Auto (default) follows
the hint recorded in the .cmf. Surfaced in /admin/api/cmf per-server status.

Cargo.lock

@@ -285,7 +285,7 @@ checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
 
 [[package]]
 name = "cortiq-gateway"
-version = "0.2.18"
+version = "0.2.19"
 dependencies = [
  "anyhow",
  "arc-swap",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cortiq-gateway"
-version = "0.2.18"
+version = "0.2.19"
 edition = "2021"
 rust-version = "1.75"
 license = "Apache-2.0"

src/admin/mod.rs

@@ -1009,6 +1009,7 @@ async fn cmf_status(State(state): State<SharedState>) -> ApiResult {
                 "port": s.port,
                 "threads": s.threads,
                 "gpu": s.gpu,
+                "o1": s.o1,
                 "running": live.map(|l| l.running).unwrap_or(false),
                 "healthy": live.map(|l| l.healthy).unwrap_or(false),
                 "last_error": live.and_then(|l| l.last_error.clone()),

src/cmf_runtime.rs

@@ -295,6 +295,11 @@ fn spawn_serve(server: &CmfServer, cfg: &CmfCfg) -> Result<Child, String> {
     if server.gpu {
         cmd.env("CMF_GPU", "1");
     }
+    // Serve-time O(1) attention override: "off" forces exact attention (undoes a
+    // too-aggressive --o1 conversion hint without re-converting the file).
+    if !server.o1.trim().is_empty() {
+        cmd.env("CMF_O1", server.o1.trim());
+    }
     cmd.stdout(Stdio::null())
         .stderr(Stdio::null())
         .kill_on_drop(true)

src/config.rs

@@ -83,6 +83,12 @@ pub struct CmfServer {
     /// Offload matvecs to the GPU (CMF_GPU=1).
     #[serde(default)]
     pub gpu: bool,
+    /// Serve-time O(1) attention override (exported as CMF_O1). Empty = follow
+    /// the hint recorded in the .cmf; "off" = force exact attention (fixes a
+    /// too-aggressive conversion without re-converting); "all"/"deepN"/"i,j,k"
+    /// = force-enable on those layers.
+    #[serde(default)]
+    pub o1: String,
 }
 
 /// CMF-format model factory: import a HuggingFace model, convert it to a
@@ -172,6 +178,7 @@ impl CmfCfg {
                 port: self.local_port,
                 threads: self.threads,
                 gpu: self.gpu,
+                o1: String::new(),
             }];
         }
         Vec::new()

web/js/i18n/en.js

@@ -264,7 +264,9 @@ export default {
   "settings.cmf.pick": "select a model",
   "settings.cmf.models": "Local models",
   "settings.cmf.modelsHelp": "Each row is one model the gateway runs (its own cortiq serve on its own port) and registers in the pool. Add as many as you have .cmf files and free ports.",
-  "settings.cmf.rowHint": "id · .cmf file · port · threads · GPU. Changes apply on restart; deleting a row stops that server.",
+  "settings.cmf.rowHint": "id · .cmf file · port · threads · GPU · O(1). Changes apply on restart; deleting a row stops that server.",
+  "settings.cmf.o1Auto": "O(1): auto",
+  "settings.cmf.o1Help": "O(1) attention at serve time: auto — follow the hint recorded in the .cmf; off — force exact attention (fixes incoherent output after a too-aggressive --o1 conversion, no re-convert needed); all/deepN — force-enable on those layers.",
   "settings.cmf.addServer": "+ Add local model",
   "settings.cmf.checkPort": "Check port",
   "settings.cmf.portFree": "port is free",

web/js/i18n/ru.js

@@ -263,7 +263,9 @@ export default {
   "settings.cmf.pick": "выберите модель",
   "settings.cmf.models": "Локальные модели",
   "settings.cmf.modelsHelp": "Каждая строка — отдельная модель: шлюз поднимает свой cortiq serve на своём порту и регистрирует её в пуле. Добавляйте столько, сколько есть .cmf-файлов и свободных портов.",
-  "settings.cmf.rowHint": "id · файл .cmf · порт · потоки · GPU. Изменения применяются после перезапуска; удаление строки останавливает сервер.",
+  "settings.cmf.rowHint": "id · файл .cmf · порт · потоки · GPU · O(1). Изменения применяются после перезапуска; удаление строки останавливает сервер.",
+  "settings.cmf.o1Auto": "O(1): авто",
+  "settings.cmf.o1Help": "O(1)-внимание при запуске: авто — использовать подсказку из .cmf; off — принудительно точное внимание (лечит бессвязный вывод после слишком агрессивной --o1 конвертации, без переконвертации); all/deepN — принудительно включить на этих слоях.",
   "settings.cmf.addServer": "+ Добавить модель",
   "settings.cmf.checkPort": "Проверить порт",
   "settings.cmf.portFree": "порт свободен",

web/js/views/settings.js

@@ -234,6 +234,12 @@ export async function renderSettings() {
     const portIn = h("input", { type: "number", value: sv.port ?? 8090, style: "width:82px" });
     const threadsIn = h("input", { type: "number", min: 0, value: sv.threads ?? 8, style: "width:60px", title: t("settings.cmf.threadsHelp") });
     const gpuIn = h("input", { type: "checkbox", checked: sv.gpu ? true : null, title: t("settings.cmf.gpuHelp") });
+    // serve-time O(1) override: off cures a too-aggressive --o1 conversion
+    const o1Opts = ["", "off", "all", "deep4", "deep8", "deep12"];
+    const o1In = h("select", { style: "width:110px", title: t("settings.cmf.o1Help") },
+      ...o1Opts.map((v) => h("option", { value: v, selected: (sv.o1 || "") === v ? true : null },
+        v === "" ? t("settings.cmf.o1Auto") : v)));
+    if (sv.o1 && !o1Opts.includes(sv.o1)) o1In.appendChild(h("option", { value: sv.o1, selected: true }, sv.o1));
     const checkBtn = h("button", { class: "btn small", type: "button" }, t("settings.cmf.checkPort"));
     const removeBtn = h("button", { class: "btn small danger", type: "button", title: t("common.delete") }, "✕");
     const st = h("span", { class: "small" });
@@ -258,8 +264,8 @@ export async function renderSettings() {
       checkBtn.disabled = false;
     });
     const row = h("div", { class: "flex wrap", style: "gap:6px;align-items:center;margin:6px 0" },
-      idIn, modelSel, portIn, checkBtn, threadsIn, h("label", { class: "flex", style: "gap:4px" }, gpuIn, "GPU"), removeBtn, st);
-    const entry = { node: row, read: () => ({ id: idIn.value.trim() || autoId(modelSel.value), model: modelSel.value.trim(), port: parseInt(portIn.value) || 8090, threads: parseInt(threadsIn.value) || 0, gpu: gpuIn.checked }) };
+      idIn, modelSel, portIn, checkBtn, threadsIn, h("label", { class: "flex", style: "gap:4px" }, gpuIn, "GPU"), o1In, removeBtn, st);
+    const entry = { node: row, read: () => ({ id: idIn.value.trim() || autoId(modelSel.value), model: modelSel.value.trim(), port: parseInt(portIn.value) || 8090, threads: parseInt(threadsIn.value) || 0, gpu: gpuIn.checked, o1: o1In.value }) };
     removeBtn.addEventListener("click", () => { const i = serverRows.indexOf(entry); if (i >= 0) serverRows.splice(i, 1); row.remove(); });
     serverRows.push(entry);
     serversWrap.appendChild(row);