Fix logit bias for DeepSeek on OpenRouter

96143177ced2bb8c283af0510268f6582599950d

Cohee <18619528+Cohee1207@users.noreply.github.com>

src/endpoints/backends/chat-completions.js+10 -0 Viewed

     getTiktokenTokenizer,
     sentencepieceTokenizers,
     TEXT_COMPLETION_MODELS,
+    webTokenizers,
+    getWebTokenizer,
 } from '../tokenizers.js';
 const API_OPENAI = 'https://api.openai.com/v1';

                 return response.send({});
+            }
             encodeFunction = (text) => new Uint32Array(instance.encodeIds(text));
+        } else if (webTokenizers.includes(model)) {
+            const tokenizer = getWebTokenizer(model);
+            const instance = await tokenizer?.get();
+            if (!instance) {
+                console.warn('Tokenizer not initialized:', model);
+                return response.send({});
+            }
+            encodeFunction = (text) => new Uint32Array(instance.encode(text));
         } else {
             const tokenizer = getTiktokenTokenizer(model);
             encodeFunction = (tokenizer.encode.bind(tokenizer));

src/endpoints/tokenizers.js+42 -0 Viewed

     'jamba',
 ];
+export const webTokenizers = [
+    'claude',
+    'llama3',
+    'command-r',
+    'qwen2',
+    'nemo',
+    'deepseek',
+];
 /**
  * Gets the Sentencepiece tokenizer by the model name.
  * @param {string} model Sentencepiece model name

+}
 /**
+ * Gets the Web tokenizer by the model name.
+ * @param {string} model Web tokenizer model name
+ * @returns {WebTokenizer|null} Web tokenizer
+ */
+export function getWebTokenizer(model) {
+    if (model.includes('llama3')) {
+        return llama3_tokenizer;
+    }
+    if (model.includes('claude')) {
+        return claude_tokenizer;
+    }
+    if (model.includes('command-r')) {
+        return commandTokenizer;
+    }
+    if (model.includes('qwen2')) {
+        return qwen2Tokenizer;
+    }
+    if (model.includes('nemo')) {
+        return nemoTokenizer;
+    }
+    if (model.includes('deepseek')) {
+        return deepseekTokenizer;
+    }
+    return null;
+}
+/**
  * Counts the token ids for the given text using the Sentencepiece tokenizer.
  * @param {SentencePieceTokenizer} tokenizer Sentencepiece tokenizer
  * @param {string} text Text to tokenize