llama : add support for SmolLm pre-tokenizer (#8609)

* Adding SmolLM Pre Tokenizer

* Update convert_hf_to_gguf_update.py

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

* Update src/llama.cpp

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

* handle regex

* removed .inp and out .out ggufs

---------

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

Author: Stillerman

convert_hf_to_gguf.py

@@ -597,6 +597,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "63b97e4253352e6f357cc59ea5b583e3a680eaeaf2632188c2b952de2588485e":
             # ref: https://huggingface.co/mistralai/Mistral-Nemo-Base-2407
             res = "tekken"
+        if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
+            # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
+            res = "smollm"
 
         if res is None:
             logger.warning("\n")

convert_hf_to_gguf_update.py

@@ -92,6 +92,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
     {"name": "t5",             "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
+    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
 ]
 
 

include/llama.h

@@ -93,6 +93,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_VIKING         = 18,
         LLAMA_VOCAB_PRE_TYPE_JAIS           = 19,
         LLAMA_VOCAB_PRE_TYPE_TEKKEN         = 20,
+        LLAMA_VOCAB_PRE_TYPE_SMOLLM         = 21,
     };
 
     // note: these values should be synchronized with ggml_rope

src/llama.cpp

@@ -5521,6 +5521,10 @@ static void llm_load_vocab(
                 vocab.tokenizer_clean_spaces = false;
                 vocab.tokenizer_ignore_merges = true;
                 vocab.tokenizer_add_bos = true;
+            } else if (
+                tokenizer_pre == "smollm") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMOLLM;
+                vocab.tokenizer_clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -15543,6 +15547,7 @@ struct llm_tokenizer_bpe {
             case LLAMA_VOCAB_PRE_TYPE_STARCODER:
             case LLAMA_VOCAB_PRE_TYPE_REFACT:
             case LLAMA_VOCAB_PRE_TYPE_COMMAND_R:
+            case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",