Added support for Viking pre-tokenizer (ggml-org#8135)

Co-authored-by: kustaaya <[email protected]>

Author: 2 people

convert-hf-to-gguf-update.py

@@ -85,6 +85,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
     {"name": "poro-chat",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Poro-34B-chat", },
     {"name": "jina-v2-code",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
+    {"name": "viking",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B
 ]
 
 

convert-hf-to-gguf.py

@@ -487,6 +487,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
             res = "jina-v2-code"
+        if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
+            # ref: https://huggingface.co/LumiOpen/Viking-7B
+            res = "viking"
 
         if res is None:
             logger.warning("\n")

llama.cpp

@@ -5193,6 +5193,9 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "poro-chat") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
+            } else if (
+                tokenizer_pre == "viking") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_VIKING;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -14078,6 +14081,12 @@ struct llm_tokenizer_bpe {
                     " ?[^(\\s|.,!?…。，、।۔،)]+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_VIKING:
+                regex_exprs = {
+                    "\\p{N}",
+                    " ?[^(\\s|.,!?…。，、।۔،)]+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {

llama.h

@@ -88,6 +88,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
         LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
         LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
+        LLAMA_VOCAB_PRE_TYPE_VIKING         = 16,
     };
 
     // note: these values should be synchronized with ggml_rope