@khanglvm/llm-router 2.3.1 → 2.3.2

package/src/node/ollama-client.js

@@ -0,0 +1,195 @@
+/** Ollama REST API client. All exports return { ok, error?, ... } — never throw. */
+
+const DEFAULT_TIMEOUT_MS = 5_000;
+const LOAD_TIMEOUT_MS = 120_000;
+const PULL_TIMEOUT_MS = 600_000;
+
+/**
+ * @param {string} baseUrl  e.g. "http://localhost:11434"
+ * @param {string} path     must start with "/"
+ * @param {RequestInit & { timeoutMs?: number }} options
+ * @returns {Promise<{ ok: boolean, status: number, json: unknown, error: string | null }>}
+ */
+async function ollamaFetch(baseUrl, path, options = {}) {
+  const { timeoutMs = DEFAULT_TIMEOUT_MS, ...init } = options;
+  const url = baseUrl.replace(/\/+$/, "") + path;
+  let response;
+  let json = null;
+  let error = null;
+
+  try {
+    response = await fetch(url, {
+      ...init,
+      signal: init.signal ?? AbortSignal.timeout(timeoutMs)
+    });
+    const text = await response.text();
+    if (text) {
+      try {
+        json = JSON.parse(text);
+      } catch {
+        // non-JSON body — leave json null
+      }
+    }
+  } catch (err) {
+    error = err instanceof Error ? err.message : String(err);
+  }
+
+  return {
+    ok: Boolean(response?.ok),
+    status: response?.status ?? 0,
+    json,
+    error
+  };
+}
+
+/** Check whether the Ollama server is reachable. */
+export async function ollamaCheckConnection(baseUrl) {
+  const result = await ollamaFetch(baseUrl, "/");
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  return { ok: true };
+}
+
+/** List all locally available models. */
+export async function ollamaListModels(baseUrl) {
+  const result = await ollamaFetch(baseUrl, "/api/tags");
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  const raw = result.json?.models ?? [];
+  const models = raw.map((m) => ({
+    name: m.name,
+    parameterSize: m.details?.parameter_size ?? null,
+    quantizationLevel: m.details?.quantization_level ?? null,
+    sizeBytes: m.size ?? null,
+    family: m.details?.family ?? null,
+    modifiedAt: m.modified_at ?? null,
+    contextLength: null
+  }));
+  // Enrich with contextLength from /api/show (max 5 concurrent)
+  const BATCH = 5;
+  for (let i = 0; i < models.length; i += BATCH) {
+    const batch = models.slice(i, i + BATCH);
+    const details = await Promise.all(
+      batch.map((m) => ollamaShowModel(baseUrl, m.name).catch(() => ({ ok: false })))
+    );
+    for (let j = 0; j < batch.length; j++) {
+      if (details[j]?.ok && details[j].details?.contextLength) {
+        batch[j].contextLength = details[j].details.contextLength;
+      }
+    }
+  }
+  return { ok: true, models };
+}
+
+/** Show detailed info for a specific model. */
+export async function ollamaShowModel(baseUrl, modelName) {
+  const result = await ollamaFetch(baseUrl, "/api/show", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model: modelName, verbose: false })
+  });
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  const body = result.json ?? {};
+  const details = {
+    contextLength: body.model_info?.["llm.context_length"] ?? null,
+    parameterSize: body.details?.parameter_size ?? null,
+    quantizationLevel: body.details?.quantization_level ?? null,
+    family: body.details?.family ?? null,
+    format: body.details?.format ?? null
+  };
+  return { ok: true, details };
+}
+
+/** List currently running (loaded) models. */
+export async function ollamaListRunning(baseUrl) {
+  const result = await ollamaFetch(baseUrl, "/api/ps");
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  const raw = result.json?.models ?? [];
+  const PINNED_SENTINEL = "0001-01-01T00:00:00Z";
+  const models = raw.map((m) => ({
+    name: m.name,
+    sizeVram: m.size_vram ?? null,
+    expiresAt: m.expires_at ?? null,
+    isPinned: m.expires_at === PINNED_SENTINEL,
+    processor: m.details?.families?.[0] ?? null
+  }));
+  return { ok: true, models };
+}
+
+/**
+ * Load a model into memory. keepAlive: "24h" | "10m" | -1 (pin) | 0 (unload).
+ * Uses 120 s timeout to accommodate large models.
+ */
+export async function ollamaLoadModel(baseUrl, modelName, keepAlive = "24h") {
+  const result = await ollamaFetch(baseUrl, "/api/generate", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model: modelName, prompt: "", keep_alive: keepAlive, stream: false }),
+    timeoutMs: LOAD_TIMEOUT_MS
+  });
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  const loadDurationMs =
+    typeof result.json?.load_duration === "number"
+      ? result.json.load_duration / 1_000_000
+      : null;
+  return { ok: true, loadDurationMs };
+}
+
+/** Unload a model from memory immediately. */
+export async function ollamaUnloadModel(baseUrl, modelName) {
+  const result = await ollamaFetch(baseUrl, "/api/generate", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model: modelName, prompt: "", keep_alive: 0, stream: false }),
+    timeoutMs: LOAD_TIMEOUT_MS
+  });
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  return { ok: true, unloaded: true };
+}
+
+/** Pin a model in memory indefinitely (keep_alive = -1). */
+export async function ollamaPinModel(baseUrl, modelName) {
+  return ollamaLoadModel(baseUrl, modelName, -1);
+}
+
+/** Set a custom keep-alive duration for a loaded model (e.g. "10m", "1h"). */
+export async function ollamaSetKeepAlive(baseUrl, modelName, duration) {
+  return ollamaLoadModel(baseUrl, modelName, duration);
+}
+
+/** Pull (download) a model from the Ollama registry. 600 s timeout. */
+export async function ollamaPullModel(baseUrl, modelName) {
+  const result = await ollamaFetch(baseUrl, "/api/pull", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ name: modelName, stream: false }),
+    timeoutMs: PULL_TIMEOUT_MS
+  });
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  return { ok: true };
+}
+
+/** Delete a locally stored model. */
+export async function ollamaDeleteModel(baseUrl, modelName) {
+  const result = await ollamaFetch(baseUrl, "/api/delete", {
+    method: "DELETE",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ name: modelName })
+  });
+  if (!result.ok) {
+    return { ok: false, error: result.error ?? `HTTP ${result.status}` };
+  }
+  return { ok: true };
+}

package/src/node/ollama-hardware.js

@@ -0,0 +1,94 @@
+/**
+ * VRAM/memory estimation utilities for Ollama models.
+ * Pure functions, no side effects, no external dependencies.
+ */
+
+const QUANT_BITS = {
+  q2_k: 2.5,
+  q3_k_s: 3.0, q3_k_m: 3.5, q3_k_l: 3.5,
+  q4_0: 4.0, q4_k_s: 4.5, q4_k_m: 4.5,
+  q5_0: 5.0, q5_k_s: 5.5, q5_k_m: 5.5,
+  q6_k: 6.5, q8_0: 8.0, f16: 16.0, f32: 32.0
+};
+
+const DEFAULT_BITS = 4.5; // Q4_K_M
+const OVERHEAD_BYTES = 512 * 1024 * 1024; // 512 MB
+
+/**
+ * Parse a parameter size string like "4.3B", "70B" into numeric count.
+ * @param {string} parameterSize
+ * @returns {number|null} Parameter count as integer, or null if unparseable
+ */
+export function parseParameterSize(parameterSize) {
+  const match = String(parameterSize || '').trim().match(/^([\d.]+)([BKMGT]?)$/i);
+  if (!match) return null;
+
+  const value = parseFloat(match[1]);
+  if (!Number.isFinite(value) || value <= 0) return null;
+
+  const suffix = match[2].toUpperCase();
+  const multipliers = { B: 1e9, M: 1e6, K: 1e3, G: 1e9, T: 1e12, '': 1 };
+  const multiplier = multipliers[suffix] ?? 1;
+
+  return Math.round(value * multiplier);
+}
+
+/**
+ * Estimate VRAM required for a model at a given quantization and context length.
+ * @param {string} parameterSize - e.g. "7B", "70B"
+ * @param {string} quantLevel - e.g. "Q4_K_M", "F16"
+ * @param {number} contextLength - number of tokens in context window
+ * @returns {{ baseModelBytes: number, kvCacheBytes: number, totalBytes: number }|null}
+ */
+export function estimateModelVram(parameterSize, quantLevel, contextLength) {
+  const params = parseParameterSize(parameterSize);
+  if (params === null) return null;
+
+  const bitsPerWeight = QUANT_BITS[String(quantLevel || '').toLowerCase()] ?? DEFAULT_BITS;
+  const baseModelBytes = params * bitsPerWeight / 8;
+
+  // Empirical KV cache estimate: bytes per token scales with model size
+  const kvBytesPerToken = params * 0.00000025;
+  const kvCacheBytes = (Number(contextLength) || 0) * kvBytesPerToken;
+
+  const totalBytes = baseModelBytes + kvCacheBytes + OVERHEAD_BYTES;
+  return { baseModelBytes, kvCacheBytes, totalBytes };
+}
+
+/**
+ * Calculate max practical context given available memory.
+ * @param {string} parameterSize - e.g. "7B"
+ * @param {string} quantLevel - e.g. "Q4_K_M"
+ * @param {number} availableMemoryBytes
+ * @returns {{ maxContext: number, warningThreshold: number }}
+ */
+export function estimateMaxContext(parameterSize, quantLevel, availableMemoryBytes) {
+  const params = parseParameterSize(parameterSize);
+  if (params === null) return { maxContext: 0, warningThreshold: 0 };
+
+  const bitsPerWeight = QUANT_BITS[String(quantLevel || '').toLowerCase()] ?? DEFAULT_BITS;
+  const baseModelBytes = params * bitsPerWeight / 8;
+
+  const remaining = availableMemoryBytes - baseModelBytes - OVERHEAD_BYTES;
+  if (remaining <= 0) return { maxContext: 0, warningThreshold: 0 };
+
+  const kvBytesPerToken = params * 0.00000025;
+  const maxContext = Math.floor(remaining / kvBytesPerToken / 1024) * 1024;
+  const warningThreshold = Math.floor(maxContext * 0.85);
+
+  return { maxContext, warningThreshold };
+}
+
+/**
+ * Format a byte count as a human-readable string.
+ * @param {number} bytes
+ * @returns {string} e.g. "4.5 GB", "512.0 MB"
+ */
+export function formatBytes(bytes) {
+  const n = Number(bytes) || 0;
+  if (n >= 1024 ** 4) return `${(n / 1024 ** 4).toFixed(1)} TB`;
+  if (n >= 1024 ** 3) return `${(n / 1024 ** 3).toFixed(1)} GB`;
+  if (n >= 1024 ** 2) return `${(n / 1024 ** 2).toFixed(1)} MB`;
+  if (n >= 1024)      return `${(n / 1024).toFixed(1)} KB`;
+  return `${n} B`;
+}

package/src/node/ollama-install.js

@@ -0,0 +1,230 @@
+/**
+ * Ollama detection, installation, and server lifecycle management.
+ * All public functions return structured results — never throw.
+ */
+
+import { spawnSync, spawn } from "node:child_process";
+import { existsSync } from "node:fs";
+
+const OLLAMA_PORT_URL = "http://localhost:11434/";
+const STARTUP_WAIT_MS = 2_000;
+const HEALTH_TIMEOUT_MS = 3_000;
+
+/** @type {Record<string, string[]>} */
+const FALLBACK_PATHS = {
+  darwin: ["/usr/local/bin/ollama", "/opt/homebrew/bin/ollama"],
+  linux: ["/usr/local/bin/ollama", "/usr/bin/ollama"],
+  win32: ["C:\\Program Files\\Ollama\\ollama.exe"]
+};
+
+/**
+ * Detect if Ollama is installed on the system.
+ * @returns {{ installed: boolean, path: string, version: string }}
+ */
+export function detectOllamaInstallation() {
+  try {
+    const platform = process.platform;
+    const whichCmd = platform === "win32" ? "where" : "which";
+    const which = spawnSync(whichCmd, ["ollama"], { encoding: "utf8" });
+    let ollamaPath = which.stdout?.trim() ?? "";
+
+    if (!ollamaPath) {
+      const candidates = FALLBACK_PATHS[platform] ?? [];
+      ollamaPath = candidates.find((p) => existsSync(p)) ?? "";
+    }
+
+    if (!ollamaPath) {
+      return { installed: false, path: "", version: "" };
+    }
+
+    const ver = spawnSync("ollama", ["--version"], { encoding: "utf8" });
+    const version = ver.stdout?.trim() ?? "";
+    return { installed: true, path: ollamaPath, version };
+  } catch {
+    return { installed: false, path: "", version: "" };
+  }
+}
+
+/**
+ * Install Ollama silently per platform.
+ * @param {{ onProgress?: (event: { phase: string, message: string }) => void }} opts
+ * @returns {Promise<{ ok: boolean, version?: string, error?: string, alreadyInstalled?: boolean }>}
+ */
+export async function installOllama({ onProgress } = {}) {
+  const progress = (phase, message) => onProgress?.({ phase, message });
+
+  try {
+    progress("detecting", "Checking for existing Ollama installation...");
+    const existing = detectOllamaInstallation();
+    if (existing.installed) {
+      progress("done", "Ollama is already installed.");
+      return { ok: true, alreadyInstalled: true, version: existing.version };
+    }
+
+    const platform = process.platform;
+
+    if (platform === "win32") {
+      const msg = "Automatic install not supported on Windows. Please install from https://ollama.com/download";
+      progress("error", msg);
+      return { ok: false, error: msg };
+    }
+
+    if (platform === "darwin") {
+      return await installViaBrew({ progress });
+    }
+
+    if (platform === "linux") {
+      return await installViaScript({ progress });
+    }
+
+    const msg = `Unsupported platform: ${platform}`;
+    progress("error", msg);
+    return { ok: false, error: msg };
+  } catch (err) {
+    const error = err instanceof Error ? err.message : String(err);
+    progress("error", error);
+    return { ok: false, error };
+  }
+}
+
+/**
+ * Start Ollama server as a detached background process.
+ * @returns {Promise<{ ok: boolean, pid?: number, error?: string }>}
+ */
+export async function startOllamaServer() {
+  try {
+    const child = spawn("ollama", ["serve"], {
+      detached: true,
+      stdio: "ignore"
+    });
+    child.unref();
+    const pid = child.pid;
+
+    await new Promise((resolve) => setTimeout(resolve, STARTUP_WAIT_MS));
+
+    const running = await isOllamaRunning();
+    if (!running) {
+      return { ok: false, error: "Server did not respond after startup" };
+    }
+
+    return { ok: true, pid };
+  } catch (err) {
+    return { ok: false, error: err instanceof Error ? err.message : String(err) };
+  }
+}
+
+/**
+ * Stop the Ollama server process.
+ * @returns {{ ok: boolean }}
+ */
+export function stopOllamaServer() {
+  try {
+    if (process.platform === "win32") {
+      spawnSync("taskkill", ["/IM", "ollama.exe", "/F"]);
+    } else {
+      spawnSync("pkill", ["-x", "ollama"], { timeout: 5000 });
+    }
+    return { ok: true };
+  } catch {
+    return { ok: false };
+  }
+}
+
+/**
+ * Check if Ollama server is responding.
+ * @returns {Promise<boolean>}
+ */
+export async function isOllamaRunning() {
+  try {
+    const res = await fetch(OLLAMA_PORT_URL, {
+      signal: AbortSignal.timeout(HEALTH_TIMEOUT_MS)
+    });
+    return res.ok || res.status > 0;
+  } catch {
+    return false;
+  }
+}
+
+// -- Private helpers --
+
+async function installViaBrew({ progress }) {
+  const brew = spawnSync("which", ["brew"], { encoding: "utf8" });
+  if (!brew.stdout?.trim()) {
+    const msg =
+      "Homebrew not found. Please install Ollama manually from https://ollama.com/download";
+    progress("error", msg);
+    return { ok: false, error: msg };
+  }
+
+  progress("downloading", "Installing Ollama via Homebrew...");
+  return new Promise((resolve) => {
+    const child = spawn("brew", ["install", "ollama"], { stdio: "pipe" });
+
+    child.stdout.on("data", (d) =>
+      progress("installing", d.toString().trim())
+    );
+    child.stderr.on("data", (d) =>
+      progress("installing", d.toString().trim())
+    );
+
+    child.on("close", (code) => {
+      if (code !== 0) {
+        const error = `brew install ollama exited with code ${code}`;
+        progress("error", error);
+        return resolve({ ok: false, error });
+      }
+      progress("verifying", "Verifying installation...");
+      const result = detectOllamaInstallation();
+      if (!result.installed) {
+        const error = "Installation succeeded but ollama binary not found";
+        progress("error", error);
+        return resolve({ ok: false, error });
+      }
+      progress("done", "Ollama installed successfully.");
+      resolve({ ok: true, version: result.version });
+    });
+
+    child.on("error", (err) => {
+      progress("error", err.message);
+      resolve({ ok: false, error: err.message });
+    });
+  });
+}
+
+async function installViaScript({ progress }) {
+  progress("downloading", "Downloading Ollama install script...");
+  return new Promise((resolve) => {
+    const child = spawn("sh", ["-c", "curl -fsSL https://ollama.com/install.sh | sh"], {
+      stdio: "pipe"
+    });
+
+    child.stdout.on("data", (d) =>
+      progress("installing", d.toString().trim())
+    );
+    child.stderr.on("data", (d) =>
+      progress("installing", d.toString().trim())
+    );
+
+    child.on("close", (code) => {
+      if (code !== 0) {
+        const error = `Install script exited with code ${code}`;
+        progress("error", error);
+        return resolve({ ok: false, error });
+      }
+      progress("verifying", "Verifying installation...");
+      const result = detectOllamaInstallation();
+      if (!result.installed) {
+        const error = "Installation succeeded but ollama binary not found";
+        progress("error", error);
+        return resolve({ ok: false, error });
+      }
+      progress("done", "Ollama installed successfully.");
+      resolve({ ok: true, version: result.version });
+    });
+
+    child.on("error", (err) => {
+      progress("error", err.message);
+      resolve({ ok: false, error: err.message });
+    });
+  });
+}

package/src/node/provider-probe.js

@@ -84,6 +84,9 @@ function resolveModelsUrl(baseUrl, format) {
       return clean.replace(/\/chat\/completions$/, "/models");
     }
     if (clean.endsWith("/v1") || isVersionedApiRoot) return `${clean}/models`;
+    // Handle base URLs with a versioned segment followed by a sub-path,
+    // e.g. https://generativelanguage.googleapis.com/v1beta/openai
+    if (/\/v\d+[a-z]*\/(?!chat\b)\w+$/i.test(clean)) return `${clean}/models`;
     return `${clean}/v1/models`;
   }
 
@@ -191,8 +194,12 @@ function extractModelIds(result) {
   const ids = [];
   for (const item of body.data) {
     if (!item || typeof item !== "object") continue;
-    const id = typeof item.id === "string" ? item.id : (typeof item.name === "string" ? item.name : null);
-    if (id) ids.push(id);
+    let id = typeof item.id === "string" ? item.id : (typeof item.name === "string" ? item.name : null);
+    if (id) {
+      // Strip provider-specific prefixes (e.g., Gemini "models/gemini-*")
+      if (id.startsWith("models/")) id = id.slice(7);
+      ids.push(id);
+    }
   }
   return [...new Set(ids)];
 }
@@ -553,12 +560,14 @@ async function probeOpenAI(baseUrl, apiKey, timeoutMs, extraHeaders = {}) {
     }, timeoutMs);
     details.checks.push({ step: "chat", auth: variant.type, status: chatResult.status, error: chatResult.error || null });
 
-    if (looksOpenAI(chatResult)) {
+    const modelsLooksValid = looksOpenAI(modelsResult) && authLooksValid(modelsResult);
+
+    if (looksOpenAI(chatResult) || modelsLooksValid) {
       details.supported = true;
-      if (authLooksValid(chatResult)) {
+      if (looksOpenAI(chatResult) ? authLooksValid(chatResult) : modelsLooksValid) {
         details.working = true;
         details.auth = { type: variant.type === "x-api-key" ? "x-api-key" : "bearer" };
-        if (looksOpenAI(modelsResult) && authLooksValid(modelsResult)) {
+        if (modelsLooksValid) {
           details.models = extractModelIds(modelsResult);
         }
         return details;
@@ -799,6 +808,61 @@ function pickPreferredFormatForModel(modelId, formats, { providerPreferredFormat
   return supported[0];
 }
 
+/**
+ * Probes a list of models against an OpenAI-compatible endpoint to detect
+ * free-tier availability. Returns a map of modelId -> { freeTier, rpm }.
+ * A model is considered not-free-tier if the response contains "limit: 0"
+ * in a free-tier quota metric.
+ */
+export async function probeFreeTierModels(options) {
+  const baseUrl = String(options?.baseUrl || "").trim().replace(/\/+$/, "");
+  const apiKey = String(options?.apiKey || "").trim();
+  const modelIds = (options?.modelIds || []).map((id) => String(id || "").trim()).filter(Boolean);
+  const timeoutMs = Number.isFinite(options?.timeoutMs) ? options.timeoutMs : 6000;
+
+  if (!baseUrl || !apiKey || modelIds.length === 0) return {};
+
+  const chatUrl = `${baseUrl}/chat/completions`;
+  const headers = {
+    "Content-Type": "application/json",
+    Authorization: `Bearer ${apiKey}`
+  };
+  const result = {};
+
+  for (const modelId of modelIds) {
+    try {
+      const response = await safeFetchJson(chatUrl, {
+        method: "POST",
+        headers,
+        body: JSON.stringify({
+          model: modelId,
+          messages: [{ role: "user", content: "hi" }],
+          max_tokens: 1,
+          stream: false
+        })
+      }, timeoutMs);
+
+      const text = response.text || "";
+      const isZeroQuota = /limit:\s*0[,\s]/i.test(text) || text.includes('"limit": 0') || text.includes('"limit":0');
+      const isFreeTierQuota = text.includes("free_tier");
+
+      if (isZeroQuota && isFreeTierQuota) {
+        result[modelId] = { freeTier: false };
+      } else if (response.ok || response.status === 400 || response.status === 404) {
+        result[modelId] = { freeTier: true };
+      } else if (response.status === 429 && !isZeroQuota) {
+        result[modelId] = { freeTier: true };
+      } else {
+        result[modelId] = { freeTier: false };
+      }
+    } catch {
+      result[modelId] = { freeTier: null };
+    }
+  }
+
+  return result;
+}
+
 export async function probeProviderEndpointMatrix(options) {
   const emitProgress = makeProgressEmitter(options?.onProgress);
   const apiKey = String(options?.apiKey || "").trim();