turbollm 1.14.3 → 1.14.5

package/README.md

@@ -537,8 +537,9 @@ the engine as a swappable component.
 
 1. Compile or download any `llama-server`-compatible binary — stock
    [llama.cpp](https://github.com/ggml-org/llama.cpp), a community fork, or your own build.
-2. Point TurboLLM at the **folder** — it scans for the `llama-server` binary, runs a
-   **capability probe**, and learns exactly which flags and features that build supports.
+2. Point TurboLLM at the **folder** — or upload the build's **.zip** — and it scans for the
+   `llama-server` binary, runs a **capability probe**, and learns exactly which flags and features
+   that build supports.
    *(Optional: paste the source repo URL so TurboLLM flags when a newer build ships.)*
 3. Activate it. The load-parameter UI **adapts to that engine** — features the build doesn't
    support are hidden; ones it adds (e.g. low-bit KV cache, NextN) light up.
@@ -562,7 +563,8 @@ switch between CUDA / ROCm / Metal / SYCL / Vulkan / CPU at any time (it downloa
 you choose, LM Studio-style).
 
 **Engine types.** **llama.cpp / GGUF**, **KoboldCpp** and **llamafile** (GGUF, every OS),
-**MLX** and **MLX-VLM** (macOS), and **vLLM** (Linux + NVIDIA) are all first-class engine kinds —
+**MLX** and **MLX-VLM** (macOS), **vLLM** (Linux + NVIDIA), and **LiteRT-LM** (Google's on-device
+runtime for `.litertlm` models, CPU or GPU, incl. Android via Termux; experimental) are all first-class engine kinds —
 install from the curated catalog, pick the right one per model, and switch from a single dropdown.
 
 **Fully supervised.** Every engine runs under a real state machine: health-gated readiness,

package/dist/{chunk-SEL52AIK.js → chunk-3Z2SF2XI.js}

@@ -48,8 +48,22 @@ var RAPID_MLX_MODEL_ALIAS = "default";
 function engineModelAlias(engineKind, modelPath) {
   if (engineKind === "rapid-mlx") return RAPID_MLX_MODEL_ALIAS;
   if (engineKind === "mlx-vlm") return modelPath ?? null;
+  if (engineKind === "litert-lm") return modelPath ?? null;
   return engineKind === "mlx" || engineKind === "vllm" || engineKind === "sglang" ? ENGINE_MODEL_ALIAS : null;
 }
+function applyEngineTokenLimit(engineKind, body) {
+  if (engineKind !== "litert-lm" || !("max_tokens" in body)) return body;
+  const b = body;
+  if (b.max_completion_tokens != null && b.max_tokens != null) {
+    const capped = Number(b.max_tokens);
+    const explicit = Number(b.max_completion_tokens);
+    if (Number.isFinite(capped) && Number.isFinite(explicit)) b.max_completion_tokens = Math.min(capped, explicit);
+  } else if (b.max_tokens != null) {
+    b.max_completion_tokens = b.max_tokens;
+  }
+  delete b.max_tokens;
+  return body;
+}
 
 // src/link/link-proxy.ts
 var FACADE_PREFIX = "/api/link";
@@ -218,17 +232,24 @@ function resolveChatUpstream(d, requestedModel) {
   if (!target) {
     return { ok: false, status: 409, code: "model_not_loaded", message: "Engine not running." };
   }
+  const engineKind = d.registry.active()?.kind ?? "";
   return {
     ok: true,
     upstream: {
-      modelField: engineModelAlias(d.registry.active()?.kind ?? "", d.manager.currentOpts()?.modelPath) ?? ms.model.key,
+      modelField: engineModelAlias(engineKind, d.manager.currentOpts()?.modelPath) ?? ms.model.key,
       modelName: ms.model.name,
       ctxMax: ms.model.ctx ?? DEFAULT_CTX,
-      target
+      target,
+      engineKind
     }
   };
 }
+function shapeForLocalEngine(upstream, body) {
+  if (upstream.remote || upstream.engineKind !== "litert-lm" || !body || typeof body !== "object" || Array.isArray(body)) return body;
+  return applyEngineTokenLimit(upstream.engineKind, { ...body });
+}
 function callChatUpstream(upstream, body, signal, fetchImpl = fetch, d) {
+  body = shapeForLocalEngine(upstream, body);
   const payload = JSON.stringify(body);
   const call = upstream.remote ? (() => {
     const headers = linkHeaders(upstream.remote);
@@ -1541,7 +1562,7 @@ var BENCH_RATE_LIMIT_MS = 60 * 60 * 1e3;
 // src/models/scanner.ts
 import { existsSync as existsSync8, lstatSync as lstatSync2, readdirSync as readdirSync5, readFileSync as readFileSync8, realpathSync as realpathSync2, rmSync as rmSync4, writeFileSync as writeFileSync6 } from "fs";
 import { readdir, realpath, stat } from "fs/promises";
-import { basename as basename3, dirname as dirname4, isAbsolute, join as join12, relative, resolve as resolve2, sep as sep2 } from "path";
+import { basename as basename4, dirname as dirname4, isAbsolute, join as join12, relative, resolve as resolve2, sep as sep2 } from "path";
 
 // src/gguf/gguf.ts
 import { closeSync, openSync, readSync } from "fs";
@@ -1550,9 +1571,37 @@ import { closeSync, openSync, readSync } from "fs";
 import { existsSync as existsSync7, lstatSync, readFileSync as readFileSync7 } from "fs";
 import { basename as basename2, join as join11 } from "path";
 
+// src/models/litertlm.ts
+import { open } from "fs/promises";
+import { basename as basename3 } from "path";
+
 // src/models/profile.ts
 var ROCM_APU_NOMMAP_MIN_BYTES = 30 * 1024 ** 3;
 
+// src/engines/litert-lm.ts
+import { createHash as createHash2 } from "crypto";
+import { existsSync as existsSync10, mkdirSync as mkdirSync7, readdirSync as readdirSync6, rmSync as rmSync6, statSync as statSync2, writeFileSync as writeFileSync7 } from "fs";
+import { execFile as execFile3 } from "child_process";
+import { dirname as dirname5, join as join14 } from "path";
+import { promisify as promisify3 } from "util";
+
+// src/engines/mlx.ts
+import { existsSync as existsSync9, mkdirSync as mkdirSync6, rmSync as rmSync5 } from "fs";
+import { execFile as execFile2 } from "child_process";
+import { join as join13 } from "path";
+import { promisify as promisify2 } from "util";
+var execFileP2 = promisify2(execFile2);
+var uvBinName = process.platform === "win32" ? "uv.exe" : "uv";
+
+// src/engines/litert-lm.ts
+var execFileP3 = promisify3(execFile3);
+var NATIVE_LOAD_PROBE = [
+  "import importlib, litert_lm",
+  "ffi = importlib.import_module('litert_lm._ffi')",
+  "load = getattr(ffi, '_get_lib', None)",
+  "load() if callable(load) else None"
+].join("\n");
+
 // src/sysinfo/sysinfo.ts
 import { execFileSync as execFileSync3 } from "child_process";
 import os from "os";
@@ -1619,7 +1668,7 @@ async function autoTitleFromConversation(d, convId) {
 }
 
 // src/code/code-session.ts
-import { join as join13, resolve as resolve3 } from "path";
+import { join as join15, resolve as resolve3 } from "path";
 import { readFile as readFile3 } from "fs/promises";
 import { Type, Unsafe } from "typebox";
 import {
@@ -2441,7 +2490,7 @@ async function runCodeSession(params) {
     console.warn(`[code-session] history seed for session ${sessionId} produced ${sessionManager.getEntries().length}/${expectedEntries} expected entries \u2014 falling back to a plain-text history prefix on this turn's prompt`);
     task = `${HISTORY_FALLBACK_PREFIX}${renderHistoryFallbackText(summaryText, priorMessages)}${HISTORY_FALLBACK_SUFFIX}${rawTask}`;
   }
-  const agentDir = join13(d.store.dir(), "pi-agent");
+  const agentDir = join15(d.store.dir(), "pi-agent");
   modelRegistry.registerProvider("local", {
     ...upstream.provider,
     api: "openai-completions",
@@ -2568,7 +2617,7 @@ async function runCodeSession(params) {
         await sink({ event: "tool_call", data });
       });
     };
-    const skillAgentDir = join13(d.store.dir(), "pi-agent");
+    const skillAgentDir = join15(d.store.dir(), "pi-agent");
     const skillResourceLoader = new DefaultResourceLoader({
       cwd: repoRoot,
       agentDir: skillAgentDir,
@@ -2709,7 +2758,7 @@ async function runCodeSession(params) {
         await sink({ event: "tool_call", data });
       });
     };
-    const subAgentDir = join13(d.store.dir(), "pi-agent");
+    const subAgentDir = join15(d.store.dir(), "pi-agent");
     const subResourceLoader = new DefaultResourceLoader({
       cwd: repoRoot,
       agentDir: subAgentDir,
@@ -3364,7 +3413,7 @@ async function compactCodeSession(params) {
       return payload;
     });
   };
-  const compactAgentDir = join13(d.store.dir(), "pi-agent");
+  const compactAgentDir = join15(d.store.dir(), "pi-agent");
   const compactResourceLoader = new DefaultResourceLoader({
     cwd: repoRoot,
     agentDir: compactAgentDir,
@@ -3463,7 +3512,7 @@ async function lookbackPreCompactionHistory(params) {
       return payload;
     });
   };
-  const lookbackAgentDir = join13(d.store.dir(), "pi-agent");
+  const lookbackAgentDir = join15(d.store.dir(), "pi-agent");
   const lookbackResourceLoader = new DefaultResourceLoader({
     cwd: repoRoot,
     agentDir: lookbackAgentDir,

package/dist/{chunk-PJPEPLCC.js → chunk-WXR3RE6R.js}

@@ -31,11 +31,14 @@ var GGMLC_GGUF = {
 };
 var NEEDS_PYTHON_ENGINE_LABEL = "needs MLX or vLLM";
 function formatIncompatibility(engineKind, format) {
-  const label = format === "gguf" ? "needs llama.cpp" : NEEDS_PYTHON_ENGINE_LABEL;
+  const label = format === "gguf" ? "needs llama.cpp" : format === "litertlm" ? "needs LiteRT-LM" : NEEDS_PYTHON_ENGINE_LABEL;
   return { code: "format", label, message: formatMismatchMessage(engineKind, format) };
 }
 function formatMismatchMessage(engineKind, format) {
   if (engineKind === "laya") return "The Laya engine runs only Laya models \u2014 load this one on another engine.";
+  if (format === "litertlm") return "This is a LiteRT-LM model (.litertlm) \u2014 activate the LiteRT-LM engine to load it.";
+  if (engineKind === "litert-lm")
+    return "The active engine is LiteRT-LM \u2014 pick a .litertlm model, or switch to the engine this model's format needs.";
   if (engineKind === "mlx")
     return "The active engine is MLX \u2014 pick a safetensors model, or switch to a llama.cpp engine for GGUF.";
   if (engineKind === "rapid-mlx")
@@ -61,6 +64,7 @@ function engineAcceptsFormat(engineKind, format) {
   if (engineKind === "mlx-vlm") return format === "mlx";
   if (engineKind === "vllm") return format === "mlx";
   if (engineKind === "laya") return false;
+  if (engineKind === "litert-lm") return format === "litertlm";
   return format === "gguf";
 }
 function engineRejectsAudioModel(engineKind) {
@@ -71,8 +75,22 @@ var RAPID_MLX_MODEL_ALIAS = "default";
 function engineModelAlias(engineKind, modelPath) {
   if (engineKind === "rapid-mlx") return RAPID_MLX_MODEL_ALIAS;
   if (engineKind === "mlx-vlm") return modelPath ?? null;
+  if (engineKind === "litert-lm") return modelPath ?? null;
   return engineKind === "mlx" || engineKind === "vllm" || engineKind === "sglang" ? ENGINE_MODEL_ALIAS : null;
 }
+function applyEngineTokenLimit(engineKind, body) {
+  if (engineKind !== "litert-lm" || !("max_tokens" in body)) return body;
+  const b = body;
+  if (b.max_completion_tokens != null && b.max_tokens != null) {
+    const capped = Number(b.max_tokens);
+    const explicit = Number(b.max_completion_tokens);
+    if (Number.isFinite(capped) && Number.isFinite(explicit)) b.max_completion_tokens = Math.min(capped, explicit);
+  } else if (b.max_tokens != null) {
+    b.max_completion_tokens = b.max_tokens;
+  }
+  delete b.max_tokens;
+  return body;
+}
 var VENDOR_DISPLAY = {
   nvidia: "NVIDIA",
   amd: "AMD",
@@ -136,7 +154,7 @@ async function distill(d, material) {
     const res = await fetch(`${target}/v1/chat/completions`, {
       method: "POST",
       headers: { "Content-Type": "application/json" },
-      body: JSON.stringify({
+      body: JSON.stringify(applyEngineTokenLimit(d.registry.active()?.kind ?? "", {
         model,
         messages: [
           { role: "system", content: DISTILL_SYSTEM },
@@ -152,7 +170,7 @@ Write ONE reusable skill as raw JSON \u2014 no markdown fences.` }
         chat_template_kwargs: { enable_thinking: false },
         response_format: { type: "json_object" },
         stream: false
-      }),
+      })),
       signal: AbortSignal.timeout(9e4)
     });
     if (!res.ok) return { name: null, description: null, procedure: null };
@@ -232,6 +250,7 @@ export {
   modelIncompatibility,
   ENGINE_MODEL_ALIAS,
   engineModelAlias,
+  applyEngineTokenLimit,
   evaluateVariant,
   distillFromConversation,
   distillSkillsFromFolder