chore: update vllm arguments

src/pages/llmodels/config/index.ts

@@ -242,3 +242,32 @@ export const setbackendParameters = (data: any) => {
   });
   return result;
 };
+
+export const getVllmCliArgs = (inputString: string) => {
+  const arrayOutput = inputString
+    .split(/\s*\[\s*|\s*\]\s*/g) // Split by '[' or ']', removing extra spaces
+    .filter((item) => item) // Remove empty items
+    .map((item) => item.trim()); // Trim spaces around each item
+
+  const result = arrayOutput.map((item) => {
+    const parts = item.split(' ');
+    const label = parts[0].trim(); // Remove leading dashes
+    const value = parts[0].trim();
+
+    if (parts.length === 1) {
+      return { label, value, options: [] };
+    } else if (parts.length > 1) {
+      const optionPart = parts[1];
+      if (optionPart.startsWith('{') && optionPart.endsWith('}')) {
+        const options = optionPart
+          .slice(1, -1)
+          .split(',')
+          .map((opt) => opt.trim());
+        return { label, value, options };
+      }
+    }
+    return { label, value, options: [] }; // Fallback
+  });
+
+  return result;
+};

src/pages/llmodels/config/vllm-config.ts

@@ -97,7 +97,12 @@ const options = [
   {
     label: '--tool-call-parser',
     value: '--tool-call-parser',
-    options: ['mistral', 'hermes']
+    options: ['mistral', 'hermes', 'internlm', 'llama3_json']
+  },
+  {
+    label: '--tool-parser-plugin',
+    value: '--tool-parser-plugin',
+    options: []
   },
   {
     label: '--model',
@@ -312,6 +317,7 @@ const options = [
       'qqq',
       'experts_int8',
       'neuron_quant',
+      'ipex',
       'None'
     ]
   },
@@ -365,6 +371,11 @@ const options = [
     value: '--limit-mm-per-prompt',
     options: []
   },
+  {
+    label: '--mm-processor-kwargs',
+    value: '--mm-processor-kwargs',
+    options: []
+  },
   {
     label: '--enable-lora',
     value: '--enable-lora',
@@ -430,6 +441,11 @@ const options = [
     value: '--num-scheduler-steps',
     options: []
   },
+  {
+    label: '--multi-step-stream-outputs',
+    value: '--multi-step-stream-outputs',
+    options: []
+  },
   {
     label: '--scheduler-delay-factor',
     value: '--scheduler-delay-factor',
@@ -467,6 +483,7 @@ const options = [
       'qqq',
       'experts_int8',
       'neuron_quant',
+      'ipex',
       'None'
     ]
   },
@@ -475,6 +492,11 @@ const options = [
     value: '--num-speculative-tokens',
     options: []
   },
+  {
+    label: '--speculative-disable-mqa-scorer',
+    value: '--speculative-disable-mqa-scorer',
+    options: []
+  },
   {
     label: '--speculative-draft-tensor-parallel-size',
     value: '--speculative-draft-tensor-parallel-size',
@@ -565,6 +587,11 @@ const options = [
     value: '--override-neuron-config',
     options: []
   },
+  {
+    label: '--scheduling-policy',
+    value: '--scheduling-policy',
+    options: ['fcfs', 'priority']
+  },
   {
     label: '--disable-log-requests',
     value: '--disable-log-requests',
@@ -574,6 +601,11 @@ const options = [
     label: '--max-log-len',
     value: '--max-log-len',
     options: []
+  },
+  {
+    label: '--disable-fastapi-docs',
+    value: '--disable-fastapi-docs',
+    options: []
   }
 ];