feat(import): O(1) attention options in Advanced (cortiq >= 0.2.0)

- New Advanced controls on the import page: O(1) attention spec (off / all /
  deep N / explicit layer list) plus the --o1-m / --o1-window / --o1-sink knobs
  with validated defaults (32 / 128 / 4) and explanatory hints.
- Backend passes --o1* to the native `cortiq convert`; fails fast with clear
  messages when combined with Python-only options, used on GGUF imports, or
  when the installed cortiq predates 0.2.0 (the hint needs the new runtime).
- The hint is recorded in the .cmf header (weights byte-identical); the runtime
  applies it at load. Verified end-to-end: converted a model with --o1 all and
  the runtime logged "o1 nystrom attention: 2/2 layer(s)".

Cargo.lock

@@ -285,7 +285,7 @@ checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
 
 [[package]]
 name = "cortiq-gateway"
-version = "0.2.14"
+version = "0.2.15"
 dependencies = [
  "anyhow",
  "arc-swap",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cortiq-gateway"
-version = "0.2.14"
+version = "0.2.15"
 edition = "2021"
 rust-version = "1.75"
 license = "Apache-2.0"

src/cmf_runtime.rs

@@ -238,7 +238,7 @@ fn parse_ver(s: &str) -> (u32, u32, u32) {
     )
 }
 
-fn version_lt(a: &str, b: &str) -> bool {
+pub(crate) fn version_lt(a: &str, b: &str) -> bool {
     parse_ver(a) < parse_ver(b)
 }
 

src/import.rs

@@ -71,6 +71,20 @@ pub struct ImportParams {
     pub shard_max_gb: Option<f32>,
     #[serde(default)]
     pub skip_mtp: bool,
+    /// O(1) Nyström attention hint (native converter, cortiq ≥ 0.2.0):
+    /// `all` | `deepN` | `i,j,k`. Weights pass through unchanged — the runtime
+    /// reads the hint at load. Empty/None/"off" = exact attention.
+    #[serde(default)]
+    pub o1: Option<String>,
+    /// Landmark budget for the o1 hint (validated default 32).
+    #[serde(default)]
+    pub o1_m: Option<usize>,
+    /// Exact-window width for the o1 hint (validated default 128).
+    #[serde(default)]
+    pub o1_window: Option<usize>,
+    /// Permanent exact sink keys for the o1 hint (validated default 4).
+    #[serde(default)]
+    pub o1_sink: Option<usize>,
 }
 fn default_quant() -> String {
     "Q8_2F".into()
@@ -269,6 +283,34 @@ pub fn start_import(store: Arc<JobStore>, cfg: &CmfCfg, p: ImportParams) -> Resu
             cfg.converter
         ));
     }
+    // O(1) attention hint — native-converter feature (cortiq ≥ 0.2.0). Fail fast
+    // with a clear message instead of a mid-job "unexpected argument" error.
+    let o1 =
+        p.o1.as_deref()
+            .map(str::trim)
+            .filter(|s| !s.is_empty() && *s != "off")
+            .map(str::to_string);
+    if o1.is_some() {
+        if use_python {
+            return Err(
+                "O(1) attention is supported by the native converter only — remove the \
+                 linear-core / v-bit shape / mean-bits options to use it"
+                    .into(),
+            );
+        }
+        if p.repo.to_ascii_lowercase().contains("gguf") {
+            return Err("O(1) attention is not supported for GGUF imports yet".into());
+        }
+        let ver = crate::cmf_runtime::installed_version(&cfg.cortiq_bin);
+        if let Some(v) = &ver {
+            if crate::cmf_runtime::version_lt(v, "0.2.0") {
+                return Err(format!(
+                    "O(1) attention needs cortiq ≥ 0.2.0 (installed: {v}) — update the runtime \
+                     in Settings → Local models"
+                ));
+            }
+        }
+    }
     let _ = std::fs::create_dir_all(&cfg.models_dir);
     let base = if p.name.trim().is_empty() {
         sanitize(p.repo.rsplit('/').next().unwrap_or(&p.repo))
@@ -384,6 +426,23 @@ pub fn start_import(store: Arc<JobStore>, cfg: &CmfCfg, p: ImportParams) -> Resu
             a.push("--hf-token".into());
             a.push(t.clone());
         }
+        // O(1) attention hint (weights unchanged; the runtime reads it at load)
+        if let Some(spec) = &o1 {
+            a.push("--o1".into());
+            a.push(spec.clone());
+            if let Some(m) = p.o1_m {
+                a.push("--o1-m".into());
+                a.push(m.to_string());
+            }
+            if let Some(w) = p.o1_window {
+                a.push("--o1-window".into());
+                a.push(w.to_string());
+            }
+            if let Some(s) = p.o1_sink {
+                a.push("--o1-sink".into());
+                a.push(s.to_string());
+            }
+        }
         (
             cfg.cortiq_bin.clone(),
             a,

web/js/i18n/en.js

@@ -400,4 +400,15 @@ export default {
   "import.adv.shard": "Shard max (GB)",
   "import.adv.noShard": "no sharding",
   "import.adv.skipMtp": "skip MTP head",
+  "import.adv.o1": "O(1) attention",
+  "import.adv.o1Hint": "Constant-memory streaming attention: the KV cache is released after prefill, so long contexts don't grow memory. Weights stay byte-identical — only a load-time hint is recorded; you can switch back anytime (serve with --o1 off). Validated on Qwen3-0.6B: all layers ≈ ×1.18 perplexity.",
+  "import.adv.o1Off": "off — exact attention (default)",
+  "import.adv.o1All": "all — convert every softmax layer",
+  "import.adv.o1Deep": "deep N — only the N deepest layers",
+  "import.adv.o1Custom": "custom — explicit layer list",
+  "import.adv.o1DeepN": "N deep layers",
+  "import.adv.o1Layers": "Layers (comma-separated)",
+  "import.adv.o1mHint": "landmark budget (default 32)",
+  "import.adv.o1windowHint": "exact-window width (default 128)",
+  "import.adv.o1sinkHint": "permanent exact sink keys (default 4)",
 };

web/js/i18n/ru.js

@@ -399,4 +399,15 @@ export default {
   "import.adv.shard": "Шард макс (ГБ)",
   "import.adv.noShard": "без шардинга",
   "import.adv.skipMtp": "пропустить MTP-голову",
+  "import.adv.o1": "O(1)-внимание",
+  "import.adv.o1Hint": "Потоковое внимание с постоянной памятью: KV-кэш освобождается после prefill, длинный контекст не раздувает память. Веса не меняются — записывается лишь подсказка на загрузку; вернуться можно всегда (запуск с --o1 off). Проверено на Qwen3-0.6B: все слои ≈ ×1.18 перплексии.",
+  "import.adv.o1Off": "выкл — точное внимание (по умолчанию)",
+  "import.adv.o1All": "all — конвертировать все softmax-слои",
+  "import.adv.o1Deep": "deep N — только N самых глубоких слоёв",
+  "import.adv.o1Custom": "вручную — явный список слоёв",
+  "import.adv.o1DeepN": "N глубоких слоёв",
+  "import.adv.o1Layers": "Слои (через запятую)",
+  "import.adv.o1mHint": "число опорных точек (по умолчанию 32)",
+  "import.adv.o1windowHint": "ширина точного окна (по умолчанию 128)",
+  "import.adv.o1sinkHint": "постоянные точные sink-ключи (по умолчанию 4)",
 };

web/js/views/import.js

@@ -125,6 +125,32 @@ export function renderImport() {
     const meanBitsIn = h("input", { type: "number", step: "0.1", placeholder: "5.0" });
     const shardIn = h("input", { type: "number", step: "0.5", placeholder: t("import.adv.noShard") });
     const skipMtp = h("input", { type: "checkbox" });
+    // O(1) attention (cortiq ≥ 0.2.0): constant-memory streaming attention hint,
+    // weights byte-identical — the runtime applies it at load.
+    const o1Sel = h("select", {},
+      opt("", t("import.adv.o1Off")),
+      opt("all", t("import.adv.o1All")),
+      opt("deep", t("import.adv.o1Deep")),
+      opt("custom", t("import.adv.o1Custom")));
+    const o1DeepIn = h("input", { type: "number", min: "1", placeholder: "12" });
+    const o1CustomIn = h("input", { placeholder: "0,4,8" });
+    const o1M = h("input", { type: "number", min: "1", placeholder: "32" });
+    const o1Window = h("input", { type: "number", min: "0", placeholder: "128" });
+    const o1Sink = h("input", { type: "number", min: "0", placeholder: "4" });
+    const o1SpecRow = h("div", { class: "row", style: "display:none" },
+      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")));
+    o1Sel.addEventListener("change", () => {
+      const on = !!o1Sel.value;
+      o1Knobs.style.display = on ? "" : "none";
+      o1SpecRow.style.display = o1Sel.value === "deep" || o1Sel.value === "custom" ? "" : "none";
+      o1DeepIn.parentElement.style.display = o1Sel.value === "deep" ? "" : "none";
+      o1CustomIn.parentElement.style.display = o1Sel.value === "custom" ? "" : "none";
+    });
     const advBody = h(
       "div",
       { class: "adv-body", style: "display:none" },
@@ -137,6 +163,9 @@ export function renderImport() {
       h("div", { class: "row" },
         field(t("import.adv.shard"), shardIn),
         field("MTP", h("label", { class: "check" }, skipMtp, t("import.adv.skipMtp")))),
+      field(t("import.adv.o1"), o1Sel, t("import.adv.o1Hint")),
+      o1SpecRow,
+      o1Knobs,
     );
     const advToggle = h("button", { class: "btn ghost sm" }, "⚙ " + t("import.adv.title"));
     advToggle.addEventListener("click", () => {
@@ -156,6 +185,19 @@ export function renderImport() {
       if (vbitSel.value) params.vbit_shape = vbitSel.value;
       if (meanBitsIn.value) params.mean_bits = parseFloat(meanBitsIn.value);
       if (shardIn.value) params.shard_max_gb = parseFloat(shardIn.value);
+      if (o1Sel.value) {
+        // spec: all | deepN | explicit layer list
+        params.o1 =
+          o1Sel.value === "deep" ? "deep" + (parseInt(o1DeepIn.value) || 12)
+          : o1Sel.value === "custom" ? o1CustomIn.value.trim()
+          : "all";
+        if (!params.o1) delete params.o1;
+        else {
+          if (o1M.value) params.o1_m = parseInt(o1M.value);
+          if (o1Window.value) params.o1_window = parseInt(o1Window.value);
+          if (o1Sink.value) params.o1_sink = parseInt(o1Sink.value);
+        }
+      }
       goBtn.disabled = true;
       try {
         await api.startImport(params);