Add plamo2 tokenizer

Author: mitmul

convert_hf_to_gguf.py

@@ -3459,11 +3459,15 @@ def set_vocab(self):
                     elif token_type_str == "BYTE":
                         toktypes.append(gguf.TokenType.BYTE)
                     else:
-                        toktypes.append(gguf.TokenType.NORMAL)
+                        # Check for PLaMo-2 special tokens
+                        token_str = token_data[0]
+                        if token_str.startswith("<|plamo:") and token_str.endswith("|>"):
+                            toktypes.append(gguf.TokenType.CONTROL)
+                        else:
+                            toktypes.append(gguf.TokenType.NORMAL)
 
-        # Use "llama" (SPM) tokenizer type which doesn't require merges
-        # PLaMo 2's tokenizer is more similar to SPM than GPT2
-        self.gguf_writer.add_tokenizer_model("llama")
+        # Use "plamo2" tokenizer type for PLaMo-2's custom Aho-Corasick tokenizer
+        self.gguf_writer.add_tokenizer_model("plamo2")
         self.gguf_writer.add_tokenizer_pre("default")
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_scores(scores)

include/llama.h

@@ -71,12 +71,13 @@ extern "C" {
     typedef int32_t llama_seq_id;
 
     enum llama_vocab_type {
-        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-        LLAMA_VOCAB_TYPE_UGM  = 4, // T5 tokenizer based on Unigram
-        LLAMA_VOCAB_TYPE_RWKV = 5, // RWKV tokenizer based on greedy tokenization
+        LLAMA_VOCAB_TYPE_NONE   = 0, // For models without vocab
+        LLAMA_VOCAB_TYPE_SPM    = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
+        LLAMA_VOCAB_TYPE_BPE    = 2, // GPT-2 tokenizer based on byte-level BPE
+        LLAMA_VOCAB_TYPE_WPM    = 3, // BERT tokenizer based on WordPiece
+        LLAMA_VOCAB_TYPE_UGM    = 4, // T5 tokenizer based on Unigram
+        LLAMA_VOCAB_TYPE_RWKV   = 5, // RWKV tokenizer based on greedy tokenization
+        LLAMA_VOCAB_TYPE_PLAMO2 = 6, // PLaMo-2 tokenizer based on Aho-Corasick with dynamic programming
     };
 
     // pre-tokenization types