Add Viking-7B tokenizer support

Author: akx

convert-hf-to-gguf-update.py

@@ -82,6 +82,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
     {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
+    {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", },  # Same for 13B and 33B
 ]
 
 

convert-hf-to-gguf.py

@@ -476,6 +476,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
             # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
             res = "smaug-bpe"
+        if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
+            # ref: https://huggingface.co/LumiOpen/Viking-7B
+            res = "viking"
 
         if res is None:
             logger.warning("\n")

llama.cpp

@@ -4549,9 +4549,10 @@ static void llm_load_vocab(
                     tokenizer_pre == "default") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
-                    tokenizer_pre == "llama3"   ||
-                    tokenizer_pre == "llama-v3" ||
-                    tokenizer_pre == "llama-bpe") {
+                    tokenizer_pre == "llama3"    ||
+                    tokenizer_pre == "llama-v3"  ||
+                    tokenizer_pre == "llama-bpe" ||
+                    tokenizer_pre == "viking-7b") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
             } else if (
                     tokenizer_pre == "deepseek-llm") {
@@ -12580,6 +12581,11 @@ struct llm_tokenizer_bpe {
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                         });
                         break;
+                    case LLAMA_VOCAB_PRE_TYPE_VIKING:
+                        word_collection = unicode_regex_split(text, {
+                            " ?[^(\\s|[.,!?…。，、।۔،])]+",
+                        });
+                        break;
                     default:
                         // default regex for BPE tokenization pre-processing
                         word_collection = unicode_regex_split(text, {

llama.h

@@ -86,6 +86,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
         LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
         LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
+        LLAMA_VOCAB_PRE_TYPE_VIKING         = 15,
     };
 
     // note: these values should be synchronized with ggml_rope