Upload folder using huggingface_hub

server/_core/llm.ts

@@ -171,50 +171,12 @@ const normalizeMessage = (message: Message) => {
   };
 };
 
-const normalizeToolChoice = (
-  toolChoice: ToolChoice | undefined,
-  tools: Tool[] | undefined
-): "none" | "auto" | ToolChoiceExplicit | undefined => {
-  if (!toolChoice) return undefined;
-
-  if (toolChoice === "none" || toolChoice === "auto") {
-    return toolChoice;
-  }
-
-  if (toolChoice === "required") {
-    if (!tools || tools.length === 0) {
-      throw new Error(
-        "tool_choice 'required' was provided but no tools were configured"
-      );
-    }
-
-    if (tools.length > 1) {
-      throw new Error(
-        "tool_choice 'required' needs a single tool or specify the tool name explicitly"
-      );
-    }
-
-    return {
-      type: "function",
-      function: { name: tools[0].function.name },
-    };
-  }
-
-  if ("name" in toolChoice) {
-    return {
-      type: "function",
-      function: { name: toolChoice.name },
-    };
-  }
-
-  return toolChoice;
-};
-
-const resolveApiUrl = () => {
-  // Always prefer HF Inference API if we have a token and no explicit forge URL
+const resolveApiUrl = (model: string) => {
   const hfToken = process.env.HF_TOKEN || process.env.HF_ACCESS_TOKEN;
-  if (hfToken && (!ENV.forgeApiUrl || ENV.forgeApiUrl.includes("manus.im"))) {
-    return "https://router.huggingface.co/v1/chat/completions";
+  
+  // If we have a token and it's a Hugging Face model, use the direct model endpoint
+  if (hfToken && model.includes("/")) {
+    return `https://api-inference.huggingface.co/models/${model}/v1/chat/completions`;
   }
   
   if (ENV.forgeApiUrl && ENV.forgeApiUrl.trim().length > 0) {
@@ -291,8 +253,10 @@ export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
     response_format,
   } = params;
 
+  const model = params.model || "huihui-ai/Qwen2.5-72B-Instruct-abliterated";
+
   const payload: Record<string, unknown> = {
-    model: params.model || "huihui-ai/Qwen2.5-72B-Instruct-abliterated",
+    model: model,
     messages: messages.map(normalizeMessage),
   };
 
@@ -308,7 +272,7 @@ export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
     payload.tool_choice = normalizedToolChoice;
   }
 
-  payload.max_tokens = 2048; // Reduced for better compatibility with HF Inference API
+  payload.max_tokens = 2048;
 
   const normalizedResponseFormat = normalizeResponseFormat({
     responseFormat,
@@ -321,8 +285,8 @@ export async function invokeLLM(params: InvokeParams): Promise<InvokeResult> {
     payload.response_format = normalizedResponseFormat;
   }
 
-  const apiUrl = resolveApiUrl();
-  console.log(`[LLM] Invoking ${payload.model} at ${apiUrl}`);
+  const apiUrl = resolveApiUrl(model);
+  console.log(`[LLM] Invoking ${model} at ${apiUrl}`);
 
   const response = await fetch(apiUrl, {
     method: "POST",