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 = "0.2.16"
+version = "0.2.19"
 dependencies = [
  "anyhow",
  "arc-swap",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cortiq-gateway"
-version = "0.2.16"
+version = "0.2.19"
 edition = "2021"
 rust-version = "1.75"
 license = "Apache-2.0"

src/admin/mod.rs

@@ -606,7 +606,10 @@ fn cmf_file_for_model(
     cmf: &crate::config::CmfCfg,
     m: &crate::config::ModelCfg,
 ) -> Option<std::path::PathBuf> {
+    // Any loopback URL counts as local — managed servers may sit on any port.
     let local = m.base_url == cmf.cortiq_server_url
+        || m.base_url.contains("//127.0.0.1:")
+        || m.base_url.contains("//localhost:")
         || m.base_url
             .contains(&format!("{}:{}", cmf.local_host, cmf.local_port));
     let stem = m.model.trim();
@@ -646,12 +649,24 @@ fn cmf_path_referenced(cfg: &crate::config::Config, path: &std::path::Path) -> b
             .any(|m| cmf_file_for_model(&cfg.cmf, m).as_deref() == Some(path))
 }
 
-async fn delete_model(State(state): State<SharedState>, Path(id): Path<String>) -> ApiResult {
+#[derive(Deserialize)]
+struct DeleteModelQuery {
+    /// Also remove the local `.cmf` from disk. Defaults to true (the historical
+    /// behaviour); the UI asks the user and passes it explicitly.
+    delete_file: Option<bool>,
+}
+
+async fn delete_model(
+    State(state): State<SharedState>,
+    Path(id): Path<String>,
+    Query(q): Query<DeleteModelQuery>,
+) -> ApiResult {
     let mut cfg = current_cfg(&state);
+    let delete_file = q.delete_file.unwrap_or(true);
 
     // Managed local models live in [cmf], not [[models]]. Deleting one stops its
-    // server, removes it from the managed config, and deletes its file — unless
-    // another model still points at the same file.
+    // server, removes it from the managed config, and (when asked) deletes its
+    // file — unless another model still points at the same file.
     let managed = cfg.cmf.effective_servers().into_iter().find(|s| s.id == id);
     if let Some(server) = managed {
         if !cfg.models.iter().any(|x| x.id == id) {
@@ -666,10 +681,10 @@ async fn delete_model(State(state): State<SharedState>, Path(id): Path<String>)
             let shared = cmf_path_referenced(&cfg, &path);
             state.cmf.stop_one(&id).await;
             state.reload(cfg)?;
-            let file_removed = if shared {
-                false
-            } else {
+            let file_removed = if delete_file && !shared {
                 remove_cmf_path(&path)
+            } else {
+                false
             };
             return ok(json!({ "ok": true, "file_removed": file_removed }));
         }
@@ -685,8 +700,8 @@ async fn delete_model(State(state): State<SharedState>, Path(id): Path<String>)
     let file = victim
         .as_ref()
         .and_then(|m| cmf_file_for_model(&cfg.cmf, m));
-    // Only remove the file if no remaining model still references it.
-    let file = file.filter(|p| !cmf_path_referenced(&cfg, p));
+    // Only remove the file if asked to AND no remaining model still references it.
+    let file = file.filter(|p| delete_file && !cmf_path_referenced(&cfg, p));
     state.reload(cfg)?;
     let file_removed = file.map(|p| remove_cmf_path(&p)).unwrap_or(false);
     ok(json!({ "ok": true, "file_removed": file_removed }))
@@ -994,6 +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/api.js

@@ -60,7 +60,8 @@ export const api = {
   listModels: () => j("GET", "/models"),
   createModel: (b) => j("POST", "/models", b),
   updateModel: (id, b) => j("PUT", "/models/" + encodeURIComponent(id), b),
-  deleteModel: (id) => j("DELETE", "/models/" + encodeURIComponent(id)),
+  deleteModel: (id, deleteFile = true) =>
+    j("DELETE", "/models/" + encodeURIComponent(id) + "?delete_file=" + (deleteFile ? "true" : "false")),
   probeModel: (id) => j("POST", "/models/" + encodeURIComponent(id) + "/probe"),
 
   // routing

web/js/i18n/en.js

@@ -132,6 +132,8 @@ export default {
   "models.form.secretMissing": "not set",
   "models.form.secretPlaceholder": "set or replace key (write-only)",
   "models.deleteConfirm": "Delete model \"{id}\"?",
+  "models.deleteFile": "also delete the model file from disk",
+  "models.deletedWithFile": "Deleted (file removed from disk)",
   "models.form.idOrModel": "Enter an id or a model name",
   "models.form.apiKey": "API Key",
   "models.form.apiKeyPlaceholder": "paste your API key",
@@ -262,7 +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",
@@ -410,6 +414,12 @@ export default {
   "import.adv.o1DeepN": "N deep layers",
   "import.adv.o1Layers": "Layers (comma-separated)",
   "import.adv.o1mHint": "landmark budget (default 32)",
+  "import.adv.o1mLabel": "Landmarks (--o1-m)",
+  "import.adv.o1mHelp": "How many landmark tokens summarize the whole older history when attention is approximated. More = closer to exact attention but slower and more memory; fewer = faster but rougher. 32 is the validated sweet spot — change only if you know why.",
+  "import.adv.o1windowLabel": "Exact window (--o1-window)",
+  "import.adv.o1windowHelp": "The last N tokens always get exact (non-approximated) attention. This protects quality on the text the model just wrote/read. Bigger = better local quality, more memory. Default 128.",
+  "import.adv.o1sinkLabel": "Sink keys (--o1-sink)",
+  "import.adv.o1sinkHelp": "The first N tokens of the context are kept exact forever ('attention sinks'). Models lean on these anchors — keeping them exact stabilizes long generations. Default 4; rarely needs changing.",
   "import.adv.o1windowHint": "exact-window width (default 128)",
   "import.adv.o1sinkHint": "permanent exact sink keys (default 4)",
 };

web/js/i18n/ru.js

@@ -131,6 +131,8 @@ export default {
   "models.form.secretMissing": "не задан",
   "models.form.secretPlaceholder": "задать или заменить ключ (только запись)",
   "models.deleteConfirm": "Удалить модель \"{id}\"?",
+  "models.deleteFile": "также удалить файл модели с диска",
+  "models.deletedWithFile": "Удалено (файл стёрт с диска)",
   "models.form.idOrModel": "Укажите id или имя модели",
   "models.form.apiKey": "API-ключ",
   "models.form.apiKeyPlaceholder": "вставьте API-ключ",
@@ -261,7 +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",
@@ -409,6 +413,12 @@ export default {
   "import.adv.o1DeepN": "N глубоких слоёв",
   "import.adv.o1Layers": "Слои (через запятую)",
   "import.adv.o1mHint": "число опорных точек (по умолчанию 32)",
+  "import.adv.o1mLabel": "Опорные точки (--o1-m)",
+  "import.adv.o1mHelp": "Сколько опорных токенов представляют всю старую историю при аппроксимации внимания. Больше — ближе к точному вниманию, но медленнее и больше памяти; меньше — быстрее, но грубее. 32 — проверенное значение; меняйте, только если понимаете зачем.",
+  "import.adv.o1windowLabel": "Точное окно (--o1-window)",
+  "import.adv.o1windowHelp": "Последние N токенов всегда обрабатываются точным (не аппроксимированным) вниманием — это защищает качество на только что написанном/прочитанном тексте. Больше — лучше локальное качество, больше памяти. По умолчанию 128.",
+  "import.adv.o1sinkLabel": "Sink-ключи (--o1-sink)",
+  "import.adv.o1sinkHelp": "Первые N токенов контекста навсегда остаются точными («якоря внимания»). Модели опираются на эти якоря — их точность стабилизирует длинную генерацию. По умолчанию 4; менять почти никогда не нужно.",
   "import.adv.o1windowHint": "ширина точного окна (по умолчанию 128)",
   "import.adv.o1sinkHint": "постоянные точные sink-ключи (по умолчанию 4)",
 };

web/js/views/import.js

@@ -15,8 +15,14 @@ const QUANTS = [
   ["F32", "import.q.f32"],
 ];
 
-function field(label, control, hint) {
-  return h("label", { class: "field" }, h("span", {}, label), control,
+// A small "?" badge with a hover/focus tooltip (same style as Settings).
+function help(text) {
+  return text ? h("span", { class: "help", title: text, tabindex: "0", role: "img", "aria-label": text }, "?") : null;
+}
+function field(label, control, hint, helpText) {
+  return h("label", { class: "field" },
+    h("span", {}, label, helpText ? " " : null, help(helpText)),
+    control,
     hint ? h("div", { class: "hint" }, hint) : null);
 }
 function opt(v, label, sel) {
@@ -141,9 +147,9 @@ export function renderImport() {
       field(t("import.adv.o1DeepN"), o1DeepIn),
       field(t("import.adv.o1Layers"), o1CustomIn));
     const o1Knobs = h("div", { class: "row", style: "display:none" },
-      field("--o1-m", o1M, t("import.adv.o1mHint")),
-      field("--o1-window", o1Window, t("import.adv.o1windowHint")),
-      field("--o1-sink", o1Sink, t("import.adv.o1sinkHint")));
+      field(t("import.adv.o1mLabel"), o1M, t("import.adv.o1mHint"), t("import.adv.o1mHelp")),
+      field(t("import.adv.o1windowLabel"), o1Window, t("import.adv.o1windowHint"), t("import.adv.o1windowHelp")),
+      field(t("import.adv.o1sinkLabel"), o1Sink, t("import.adv.o1sinkHint"), t("import.adv.o1sinkHelp")));
     o1Sel.addEventListener("change", () => {
       const on = !!o1Sel.value;
       o1Knobs.style.display = on ? "" : "none";

web/js/views/models.js

@@ -1,5 +1,5 @@
 // Models — pool table with add/edit modal, probe, and per-model secret entry.
-import { h, mount, modal, toast, confirmDialog, money } from "../ui.js";
+import { h, mount, modal, toast, money } from "../ui.js";
 import { t } from "../i18n.js";
 import { api } from "../api.js";
 import { appState } from "../app.js";
@@ -280,15 +280,31 @@ function row(m, meta, reload) {
           "button",
           {
             class: "btn sm danger",
-            onClick: async () => {
-              if (!(await confirmDialog(t("models.deleteConfirm", { id: m.id })))) return;
-              try {
-                await api.deleteModel(m.id);
-                toast(t("toast.deleted"), "good");
-                await reload();
-              } catch (e) {
-                toast(String(e.message || e), "bad");
-              }
+            onClick: () => {
+              // local models may have a .cmf on disk — let the user decide its fate
+              const isLocal = m.managed || /127\.0\.0\.1|localhost/.test(m.base_url || "");
+              const fileCb = h("input", { type: "checkbox", checked: isLocal ? true : null });
+              const body = h(
+                "div",
+                {},
+                h("p", {}, t("models.deleteConfirm", { id: m.id })),
+                isLocal
+                  ? h("label", { class: "check" }, fileCb, t("models.deleteFile"))
+                  : null
+              );
+              modal(t("common.delete") + ": " + m.id, body, async () => {
+                try {
+                  const r = await api.deleteModel(m.id, isLocal && fileCb.checked);
+                  toast(
+                    r.file_removed ? t("models.deletedWithFile") : t("toast.deleted"),
+                    "good"
+                  );
+                  await reload();
+                } catch (e) {
+                  toast(String(e.message || e), "bad");
+                  return false;
+                }
+              }, t("common.delete"));
             },
           },
           t("common.delete")

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);