gguf-py : export chat templates (#4125)

* gguf-py : export chat templates

* llama.cpp : escape new lines in gguf kv info prints

* gguf-py : bump version

* gguf-py : check chat_template type

* gguf-py : initialize chat_template

gguf-py/gguf/constants.py

@@ -56,20 +56,21 @@ class Keys:
         SCALING_FINETUNED    = "{arch}.rope.scaling.finetuned"
 
     class Tokenizer:
-        MODEL      = "tokenizer.ggml.model"
+        MODEL         = "tokenizer.ggml.model"
-        LIST       = "tokenizer.ggml.tokens"
+        LIST          = "tokenizer.ggml.tokens"
-        TOKEN_TYPE = "tokenizer.ggml.token_type"
+        TOKEN_TYPE    = "tokenizer.ggml.token_type"
-        SCORES     = "tokenizer.ggml.scores"
+        SCORES        = "tokenizer.ggml.scores"
-        MERGES     = "tokenizer.ggml.merges"
+        MERGES        = "tokenizer.ggml.merges"
-        BOS_ID     = "tokenizer.ggml.bos_token_id"
+        BOS_ID        = "tokenizer.ggml.bos_token_id"
-        EOS_ID     = "tokenizer.ggml.eos_token_id"
+        EOS_ID        = "tokenizer.ggml.eos_token_id"
-        UNK_ID     = "tokenizer.ggml.unknown_token_id"
+        UNK_ID        = "tokenizer.ggml.unknown_token_id"
-        SEP_ID     = "tokenizer.ggml.seperator_token_id"
+        SEP_ID        = "tokenizer.ggml.seperator_token_id"
-        PAD_ID     = "tokenizer.ggml.padding_token_id"
+        PAD_ID        = "tokenizer.ggml.padding_token_id"
-        ADD_BOS    = "tokenizer.ggml.add_bos_token"
+        ADD_BOS       = "tokenizer.ggml.add_bos_token"
-        ADD_EOS    = "tokenizer.ggml.add_eos_token"
+        ADD_EOS       = "tokenizer.ggml.add_eos_token"
-        HF_JSON    = "tokenizer.huggingface.json"
+        HF_JSON       = "tokenizer.huggingface.json"
-        RWKV       = "tokenizer.rwkv.world"
+        RWKV          = "tokenizer.rwkv.world"
+        CHAT_TEMPLATE = "tokenizer.chat_template"
 
 
 #

gguf-py/gguf/gguf_writer.py

@@ -399,6 +399,9 @@ class GGUFWriter:
     def add_add_eos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_EOS, value)
 
+    def add_chat_template(self, value: str) -> None:
+        self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/vocab.py

@@ -13,6 +13,7 @@ class SpecialVocab:
     merges: list[str]
     add_special_token: dict[str, bool]
     special_token_ids: dict[str, int]
+    chat_template: str | None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
@@ -24,6 +25,7 @@ class SpecialVocab:
         self.n_vocab = n_vocab
         self.load_merges = load_merges
         self.merges = []
+        self.chat_template = None
         if special_token_types is not None:
             self.special_token_types = special_token_types
         else:
@@ -67,6 +69,10 @@ class SpecialVocab:
             if not quiet:
                 print(f'gguf: Setting add_{typ}_token to {value}')
             add_handler(value)
+        if self.chat_template is not None:
+            if not quiet:
+                print(f'gguf: Setting chat_template to {self.chat_template}')
+            gw.add_chat_template(self.chat_template)
 
     def _load(self, path: Path) -> None:
         self._try_load_from_tokenizer_json(path)
@@ -132,6 +138,14 @@ class SpecialVocab:
             return True
         with open(tokenizer_config_file, encoding = 'utf-8') as f:
             tokenizer_config = json.load(f)
+        chat_template = tokenizer_config.get('chat_template')
+        if chat_template is None or isinstance(chat_template, str):
+            self.chat_template = chat_template
+        else:
+            print(
+                f'gguf: WARNING: Bad type for chat_template field in {tokenizer_config_file!r} - ignoring',
+                file = sys.stderr
+            )
         for typ in self.special_token_types:
             add_entry = tokenizer_config.get(f'add_{typ}_token')
             if isinstance(add_entry, bool):

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.5.3"
+version = "0.6.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

llama.cpp

@@ -1871,6 +1871,7 @@ struct llama_model_loader {
                 if (value.size() > MAX_VALUE_LEN) {
                     value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str());
                 }
+                replace_all(value, "\n", "\\n");
 
                 LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
             }