deploy: d10e28b — 更新 litellm.js

backend/src/litellm.js

@@ -1,250 +1,152 @@
 "use strict";
 
 /**
+ * LiteLLM Proxy Management Client
+ * Wraps LiteLLM admin API for dynamic model registration.
+ */
 
-- LiteLLM Proxy Management Client
-- Wraps LiteLLM’s admin API for dynamic model registration.
-  */
+const axios = require("axios");
+const { logger } = require("./logger");
 
-const axios = require(“axios”);
-const { logger } = require(”./logger”);
-
-const LITELLM_BASE_URL = process.env.LITELLM_BASE_URL || “http://litellm:4000”;
-const LITELLM_MASTER_KEY = process.env.LITELLM_MASTER_KEY || “sk-gateway-master-key”;
+const LITELLM_BASE_URL = process.env.LITELLM_BASE_URL || "http://litellm:4000";
+const LITELLM_MASTER_KEY = process.env.LITELLM_MASTER_KEY || "sk-gateway-master-key";
 
 const client = axios.create({
-baseURL: LITELLM_BASE_URL,
-timeout: 15000,
-headers: {
-Authorization: `Bearer ${LITELLM_MASTER_KEY}`,
-“Content-Type”: “application/json”,
-},
+  baseURL: LITELLM_BASE_URL,
+  timeout: 15000,
+  headers: {
+    Authorization: "Bearer " + LITELLM_MASTER_KEY,
+    "Content-Type": "application/json",
+  },
 });
 
-// ─── Model Registration ────────────────────────────────────────────────────
-
-/**
-
-- Register a new model in LiteLLM at runtime.
-- @param {object} model  - Our internal model record
-- @returns {string}       - LiteLLM internal model ID
-  */
-  async function registerModel(model) {
+async function registerModel(model) {
   const payload = {
-  model_name: model.name, // OpenAI-compatible alias exposed to callers
-  litellm_params: buildLitellmParams(model),
-  model_info: {
-  id: model.id,
-  description: model.description || “”,
-  model_type: model.modelType || “chat”,
-  },
+    model_name: model.name,
+    litellm_params: buildLitellmParams(model),
+    model_info: {
+      id: model.id,
+      description: model.description || "",
+      model_type: model.modelType || "chat",
+    },
   };
 
-logger.info(“Registering model with LiteLLM”, { modelName: model.name });
-
-const response = await client.post(”/model/new”, payload);
-
-// ─── BUG FIX #7: Ghost model on missing LiteLLM ID ──────────────────────
-//
-// ORIGINAL CODE:
-//   const litellmId = response.data?.model_info?.id || response.data?.id || model.id;
-//
-// The final `|| model.id` fallback silently stored our own UUID as
-// `litellm_id` whenever LiteLLM returned a response without a recognisable
-// ID field. This created a “ghost model” scenario:
-//
-//   1. DB stored our own UUID as `litellm_id`.
-//   2. On delete, deregisterModel() sent that UUID to LiteLLM /model/delete.
-//   3. LiteLLM couldn’t find it → returned an error we silently swallowed.
-//   4. Model removed from our DB but remained live inside LiteLLM’s in-memory
-//      router — still accepting real API traffic indefinitely.
-//
-// FIX: Extract the ID from the two documented response paths. If neither
-// yields an ID, emit a structured WARN (visible in logs/alerts) then fall
-// back to model.id only as a last resort so the creation flow is not broken.
-// The warning makes the deregistration risk explicit to operators.
-//
-// LiteLLM /model/new documented response shapes:
-//   v1.x+  → { model_info: { id: “<uuid>”, … }, model_name: “…” }
-//   older  → { id: “<uuid>”, … }
-// ─────────────────────────────────────────────────────────────────────────
-const litellmId =
-response.data?.model_info?.id ||
-response.data?.id ||
-null;
-
-if (!litellmId) {
-logger.warn(
-“[BUG#7] LiteLLM /model/new response contained no recognisable model ID. “ +
-“Falling back to internal UUID as litellm_id. “ +
-“Subsequent deregisterModel() calls for this model will likely fail silently, “ +
-“leaving a ghost model active inside LiteLLM. “ +
-“Inspect responseData below and verify your LiteLLM version.”,
-{
-modelName: model.name,
-internalId: model.id,
-responseTopLevelKeys: response.data ? Object.keys(response.data) : [],
-responseData: response.data,
-}
-);
-// Retain fallback so createModel() still returns a usable record.
-return model.id;
-}
+  logger.info("Registering model with LiteLLM", { modelName: model.name });
+
+  const response = await client.post("/model/new", payload);
+
+  const litellmId =
+    (response.data && response.data.model_info && response.data.model_info.id) ||
+    (response.data && response.data.id) ||
+    null;
+
+  if (!litellmId) {
+    logger.warn(
+      "LiteLLM /model/new response contained no recognisable model ID. " +
+        "Falling back to internal UUID. " +
+        "Deregistration may fail silently leaving a ghost model in LiteLLM.",
+      {
+        modelName: model.name,
+        internalId: model.id,
+        responseTopLevelKeys: response.data ? Object.keys(response.data) : [],
+        responseData: response.data,
+      }
+    );
+    return model.id;
+  }
 
-logger.info(“Model registered in LiteLLM”, { modelName: model.name, litellmId });
-return litellmId;
+  logger.info("Model registered in LiteLLM", { modelName: model.name, litellmId: litellmId });
+  return litellmId;
 }
 
-/**
-
-- Remove a model from LiteLLM.
-- 
-- @param {string} litellmId  - LiteLLM’s model ID (returned from registerModel)
-- 
-- ─── BUG FIX #6: /model/delete field-name version incompatibility ──────────
-- 
-- ORIGINAL CODE:
-- await client.post(”/model/delete”, { id: litellmId });
-- 
-- The request-body field name accepted by /model/delete has changed across
-- LiteLLM releases:
-- 
-- • Older versions (pre-v1.x):  { model_id: “<id>” }
-- • Current versions (v1.x+):   { id: “<id>” }
-- 
-- Sending only `id` against an older deployment produces a silent no-op:
-- LiteLLM returns HTTP 200 but ignores the request because it only reads
-- `model_id`. The same failure mode applies in reverse on newer versions.
-- 
-- FIX: Send BOTH fields in every request. LiteLLM’s Pydantic models use
-- `model_config = ConfigDict(extra="ignore")`, so unknown keys are silently
-- discarded — the payload is safe for all known versions.
-- 
-- If deletion silently fails after a future LiteLLM upgrade, verify the
-- current accepted field name via the running instance’s Swagger UI:
-- http://<litellm-host>:4000/docs  →  POST /model/delete
-- ─────────────────────────────────────────────────────────────────────────
-  */
-  async function deregisterModel(litellmId) {
+async function deregisterModel(litellmId) {
   if (!litellmId) return;
   try {
-  await client.post(”/model/delete”, {
-  id: litellmId,       // accepted by LiteLLM v1.x+
-  model_id: litellmId, // accepted by LiteLLM pre-v1.x
-  });
-  logger.info(“Model deregistered from LiteLLM”, { litellmId });
+    await client.post("/model/delete", {
+      id: litellmId,
+      model_id: litellmId,
+    });
+    logger.info("Model deregistered from LiteLLM", { litellmId: litellmId });
   } catch (err) {
-  // Model may not exist in LiteLLM (e.g. was never synced, or litellmId
-  // is our own UUID fallback from Bug Fix #7).  Log at warn so the operator
-  // is aware but the delete flow is not blocked.
-  logger.warn(“Could not deregister model from LiteLLM”, {
-  litellmId,
-  httpStatus: err.response?.status,
-  error: err.message,
-  });
-  }
-  }
-
-/**
-
-- List all models currently registered in LiteLLM.
-  */
-  async function listLitellmModels() {
-  const response = await client.get(”/model/info”);
-  return response.data?.data || response.data || [];
+    logger.warn("Could not deregister model from LiteLLM", {
+      litellmId: litellmId,
+      httpStatus: err.response && err.response.status,
+      error: err.message,
+    });
   }
+}
 
-/**
+async function listLitellmModels() {
+  const response = await client.get("/model/info");
+  return (response.data && response.data.data) || response.data || [];
+}
 
-- Update a model in LiteLLM (delete + re-add since update isn’t atomic).
-  */
-  async function updateModel(oldLitellmId, model) {
+async function updateModel(oldLitellmId, model) {
   await deregisterModel(oldLitellmId);
   return registerModel(model);
-  }
-
-/**
+}
 
-- Check LiteLLM liveness.
-- 
-- Uses /health/liveliness instead of /health:
-- - /health validates all registered model upstreams. If any upstream is
-- unreachable it returns an error even though LiteLLM itself is healthy,
-- causing syncModelsToLitellmWithRetry() to retry endlessly and give up.
-- - /health/liveliness only checks that the LiteLLM process is alive, which
-- is the correct signal for “ready to accept /model/new requests”.
-- This also matches the docker-compose.yml container healthcheck target.
-  */
-  async function healthCheck() {
-  const response = await client.get(”/health/liveliness”);
+async function healthCheck() {
+  const response = await client.get("/health/liveliness");
   return response.data;
-  }
-
-/**
+}
 
-- Test a model by sending a minimal chat completion request.
-  */
-  async function testModel(modelName, options = {}) {
+async function testModel(modelName, options) {
+  if (!options) { options = {}; }
   const start = Date.now();
   const messages =
-  options.messages && options.messages.length > 0
-  ? options.messages
-  : [{ role: “user”, content: options.prompt || “Say ‘OK’ in one word.” }];
+    options.messages && options.messages.length > 0
+      ? options.messages
+      : [{ role: "user", content: options.prompt || "Say OK in one word." }];
   try {
-  const response = await client.post(
-  “/v1/chat/completions”,
-  {
-  model: modelName,
-  messages,
-  max_tokens: 256,
-  stream: false,
-  },
-  { timeout: 30000 }
-  );
-  return {
-  success: true,
-  latencyMs: Date.now() - start,
-  response: response.data,
-  };
+    const response = await client.post(
+      "/v1/chat/completions",
+      {
+        model: modelName,
+        messages: messages,
+        max_tokens: 256,
+        stream: false,
+      },
+      { timeout: 30000 }
+    );
+    return {
+      success: true,
+      latencyMs: Date.now() - start,
+      response: response.data,
+    };
   } catch (err) {
-  return {
-  success: false,
-  latencyMs: Date.now() - start,
-  error: err.response?.data || err.message,
-  };
+    return {
+      success: false,
+      latencyMs: Date.now() - start,
+      error: (err.response && err.response.data) || err.message,
+    };
   }
-  }
-
-// ─── Helpers ──────────────────────────────────────────────────────────────
+}
 
 function buildLitellmParams(model) {
-const params = {
-model: model.litellmModel,
-};
-
-// Note: original code also checked `model._apiBase` which is never set
-// anywhere in the codebase — that dead reference has been removed.
-if (model.apiBase) {
-params.api_base = model.apiBase;
-}
+  const params = {
+    model: model.litellmModel,
+  };
 
-// Only include api_key if provided and non-empty
-const key = model._apiKey || model.apiKey;
-if (key && key !== “••••••••” && key.trim() !== “”) {
-params.api_key = key.trim();
-} else {
-// LiteLLM requires some api_key for most providers; use placeholder
-params.api_key = “none”;
-}
+  if (model.apiBase) {
+    params.api_base = model.apiBase;
+  }
+
+  const key = model._apiKey || model.apiKey;
+  if (key && key !== "\u2022\u2022\u2022\u2022\u2022\u2022\u2022\u2022" && key.trim() !== "") {
+    params.api_key = key.trim();
+  } else {
+    params.api_key = "none";
+  }
 
-return params;
+  return params;
 }
 
 module.exports = {
-registerModel,
-deregisterModel,
-listLitellmModels,
-updateModel,
-healthCheck,
-testModel,
+  registerModel: registerModel,
+  deregisterModel: deregisterModel,
+  listLitellmModels: listLitellmModels,
+  updateModel: updateModel,
+  healthCheck: healthCheck,
+  testModel: testModel,
 };