gguf : deprecate old FIM token KVs

ggerganov · Oct 11, 2024 · 3681540 · 3681540
1 parent 3ae8670
commit 3681540
Show file tree

Hide file tree

Showing 3 changed files with 36 additions and 14 deletions.
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
@@ -152,6 +152,8 @@ class Tokenizer:
         MERGES               = "tokenizer.ggml.merges"
         BOS_ID               = "tokenizer.ggml.bos_token_id"
         EOS_ID               = "tokenizer.ggml.eos_token_id"
+        EOT_ID               = "tokenizer.ggml.eot_token_id"
+        EOM_ID               = "tokenizer.ggml.eom_token_id"
         UNK_ID               = "tokenizer.ggml.unknown_token_id"
         SEP_ID               = "tokenizer.ggml.seperator_token_id"
         PAD_ID               = "tokenizer.ggml.padding_token_id"
@@ -168,11 +170,16 @@ class Tokenizer:
         CHAT_TEMPLATE_N      = "tokenizer.chat_template.{name}"
         CHAT_TEMPLATES       = "tokenizer.chat_templates"
         # FIM/Infill special tokens constants
+        FIM_PRE_ID           = "tokenizer.ggml.fim_pre_token_id"
+        FIM_SUF_ID           = "tokenizer.ggml.fim_suf_token_id"
+        FIM_MID_ID           = "tokenizer.ggml.fim_mid_token_id"
+        FIM_PAD_ID           = "tokenizer.ggml.fim_pad_token_id"
+        FIM_REP_ID           = "tokenizer.ggml.fim_rep_token_id"
+        FIM_SEP_ID           = "tokenizer.ggml.fim_sep_token_id"
+        # deprecated:
         PREFIX_ID            = "tokenizer.ggml.prefix_token_id"
         SUFFIX_ID            = "tokenizer.ggml.suffix_token_id"
         MIDDLE_ID            = "tokenizer.ggml.middle_token_id"
-        EOT_ID               = "tokenizer.ggml.eot_token_id"
-        EOM_ID               = "tokenizer.ggml.eom_token_id"
 
     class Adapter:
         TYPE       = "adapter.type"
@@ -1579,15 +1586,24 @@ def get_type(val: Any) -> GGUFValueType:
 KEY_TOKENIZER_MERGES     = Keys.Tokenizer.MERGES
 KEY_TOKENIZER_BOS_ID     = Keys.Tokenizer.BOS_ID
 KEY_TOKENIZER_EOS_ID     = Keys.Tokenizer.EOS_ID
+KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
+KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
 KEY_TOKENIZER_UNK_ID     = Keys.Tokenizer.UNK_ID
 KEY_TOKENIZER_SEP_ID     = Keys.Tokenizer.SEP_ID
 KEY_TOKENIZER_PAD_ID     = Keys.Tokenizer.PAD_ID
 KEY_TOKENIZER_CLS_ID     = Keys.Tokenizer.CLS_ID
 KEY_TOKENIZER_MASK_ID    = Keys.Tokenizer.MASK_ID
 KEY_TOKENIZER_HF_JSON    = Keys.Tokenizer.HF_JSON
 KEY_TOKENIZER_RWKV       = Keys.Tokenizer.RWKV
-KEY_TOKENIZER_PRIFIX_ID  = Keys.Tokenizer.PREFIX_ID
+
+KEY_TOKENIZER_FIM_PRE_ID = Keys.Tokenizer.FIM_PRE_ID
+KEY_TOKENIZER_FIM_SUF_ID = Keys.Tokenizer.FIM_SUF_ID
+KEY_TOKENIZER_FIM_MID_ID = Keys.Tokenizer.FIM_MID_ID
+KEY_TOKENIZER_FIM_PAD_ID = Keys.Tokenizer.FIM_PAD_ID
+KEY_TOKENIZER_FIM_REP_ID = Keys.Tokenizer.FIM_REP_ID
+KEY_TOKENIZER_FIM_SEP_ID = Keys.Tokenizer.FIM_SEP_ID
+
+# deprecated
+KEY_TOKENIZER_PREFIX_ID  = Keys.Tokenizer.PREFIX_ID
 KEY_TOKENIZER_SUFFIX_ID  = Keys.Tokenizer.SUFFIX_ID
 KEY_TOKENIZER_MIDDLE_ID  = Keys.Tokenizer.MIDDLE_ID
-KEY_TOKENIZER_EOT_ID     = Keys.Tokenizer.EOT_ID
-KEY_TOKENIZER_EOM_ID     = Keys.Tokenizer.EOM_ID
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
@@ -843,15 +843,6 @@ def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
 
         self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value)
 
-    def add_prefix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.PREFIX_ID, id)
-
-    def add_suffix_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.SUFFIX_ID, id)
-
-    def add_middle_token_id(self, id: int) -> None:
-        self.add_uint32(Keys.Tokenizer.MIDDLE_ID, id)
-
     def add_eot_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOT_ID, id)
 

diff --git a/src/llama.cpp b/src/llama.cpp
@@ -368,6 +368,11 @@ enum llm_kv {
 
     LLM_KV_ADAPTER_TYPE,
     LLM_KV_ADAPTER_LORA_ALPHA,
+
+    // deprecated:
+    LLM_KV_TOKENIZER_PREFIX_ID,
+    LLM_KV_TOKENIZER_SUFFIX_ID,
+    LLM_KV_TOKENIZER_MIDDLE_ID,
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
@@ -479,6 +484,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
 
     { LLM_KV_ADAPTER_TYPE,                     "adapter.type"       },
     { LLM_KV_ADAPTER_LORA_ALPHA,               "adapter.lora.alpha" },
+
+    // deprecated
+    { LLM_KV_TOKENIZER_PREFIX_ID,              "tokenizer.ggml.prefix_token_id" },
+    { LLM_KV_TOKENIZER_SUFFIX_ID,              "tokenizer.ggml.suffix_token_id" },
+    { LLM_KV_TOKENIZER_MIDDLE_ID,              "tokenizer.ggml.middle_token_id" },
 };
 
 struct LLM_KV {
@@ -6533,6 +6543,11 @@ static void llm_load_vocab(
             { LLM_KV_TOKENIZER_FIM_PAD_ID, vocab.special_fim_pad_id },
             { LLM_KV_TOKENIZER_FIM_REP_ID, vocab.special_fim_rep_id },
             { LLM_KV_TOKENIZER_FIM_SEP_ID, vocab.special_fim_sep_id },
+
+            // deprecated
+            { LLM_KV_TOKENIZER_PREFIX_ID, vocab.special_fim_pre_id },
+            { LLM_KV_TOKENIZER_SUFFIX_ID, vocab.special_fim_suf_id },
+            { LLM_KV_TOKENIZER_MIDDLE_ID, vocab.special_fim_mid_id },
         };
 
         for (const auto & it : special_token_types) {