gguf : make token scores and types optional (#3347)

convert-falcon-hf-to-gguf.py

@@ -133,8 +133,6 @@ gguf_writer.add_file_type(ftype)
 print("gguf: get tokenizer metadata")
 
 tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -177,12 +175,8 @@ for i in range(vocab_size):
         text = bytearray(pad_token)
 
     tokens.append(text)
-    scores.append(0.0)                      # dymmy
-    toktypes.append(gguf.TokenType.NORMAL)  # dummy
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)

convert-starcoder-hf-to-gguf.py

@@ -117,8 +117,6 @@ gguf_writer.add_file_type(ftype)
 print("gguf: get tokenizer metadata")
 
 tokens: list[bytearray] = []
-scores: list[float] = []
-toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -161,12 +159,8 @@ for i in range(vocab_size):
         text = bytearray(pad_token)
 
     tokens.append(text)
-    scores.append(0.0)                      # dymmy
-    toktypes.append(gguf.TokenType.NORMAL)  # dummy
 
 gguf_writer.add_token_list(tokens)
-gguf_writer.add_token_scores(scores)
-gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)

llama.cpp

@@ -1931,20 +1931,18 @@ static void llm_load_vocab(
         throw std::runtime_error("cannot find tokenizer vocab in model file\n");
     }
 
+    const float * scores = nullptr;
     const int score_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_SCORES).c_str());
-    if (score_idx == -1) {
-        throw std::runtime_error("cannot find tokenizer scores in model file\n");
+    if (score_idx != -1) {
+        scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
     }
 
-    const float * scores = (const float * ) gguf_get_arr_data(ctx, score_idx);
-
+    const int * toktypes = nullptr;
     const int toktype_idx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_TOKEN_TYPE).c_str());
-    if (toktype_idx == -1) {
-        throw std::runtime_error("cannot find token type list in GGUF file\n");
+    if (toktype_idx != -1) {
+        toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
     }
 
-    const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
-
     // determine vocab type
     {
         std::string tokenizer_name;
@@ -2012,8 +2010,8 @@ static void llm_load_vocab(
 
         auto & token_data = vocab.id_to_token[i];
         token_data.text  = std::move(word);
-        token_data.score = scores[i];
-        token_data.type  = (llama_token_type) toktypes[i];
+        token_data.score = scores ? scores[i] : 0.0f;
+        token_data.type  = toktypes ? (llama_token_type) toktypes[i] : LLAMA_TOKEN_TYPE_NORMAL;
     }
 
     // determine the newline token: LLaMA "<0x0A>" == 10 == '\n', Falcon 193 == '\n'