Special handling of Seed Coder FIM tokens (#585)

* Special handling of Seed Coder FIM tokens

* vocab: Add Seed Coder pretokenizer

* Formatting fix

* Update llama.h

5 files changed, 27 insertions, 0 deletions

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 1754e486..33be63fa 100644
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -636,6 +636,9 @@ class Model:
         if chkhsh == "877081d19cf6996e2c4ff0e1236341e9b7bde288f5311a56a937f0afbbb3aeb5":
             # ref: https://huggingface.co/deepseek-ai/DeepSeek-V3
             res = "deepseek-v3"
+        if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
+            # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
+            res = "seed-coder"
 
         if res is None:
             logger.warning("\n")
@@ -1520,6 +1523,17 @@ class LlamaModel(Model):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
+        # Apply to Seed-Coder only (and ignore otherwise)
+        if self.hparams.get("vocab_size", 32000) == 155136:
+            special_vocab = gguf.SpecialVocab(
+                self.dir_model, load_merges=False,
+                special_token_types = ['prefix', 'suffix', 'middle', 'eot']
+            )
+            special_vocab._set_special_token("prefix", 124)
+            special_vocab._set_special_token("suffix", 125)
+            special_vocab._set_special_token("middle", 126)
+            special_vocab.add_to_gguf(self.gguf_writer)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py
index 40af02f4..f2e6cc37 100755
--- a/convert_hf_to_gguf_update.py
+++ b/convert_hf_to_gguf_update.py
@@ -95,6 +95,7 @@ models = [
     {"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", },
     {"name": "deepseek-v3",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"},
+    {"name": "seed-coder",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]
 
 
diff --git a/include/llama.h b/include/llama.h
index 6c8bff95..51db3eab 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -110,6 +110,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
         LLAMA_VOCAB_PRE_TYPE_FALCON_3       = 34,
         LLAMA_VOCAB_PRE_TYPE_FALCON_E       = 35,
+        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 36, //llama.cpp lists this as 35
     };
 
     // note: these values should be synchronized with ggml_rope
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index 65ca5e38..474cbd8c 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -477,6 +477,13 @@ struct llm_tokenizer_bpe {
                     "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
                 };
                 break;
+	    case LLAMA_VOCAB_PRE_TYPE_SEED_CODER:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\r\n]+|\\s*[\r\n]+|\\s+(?!\\S)|\\s+"
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\\r\\n]+|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
diff --git a/src/llama.cpp b/src/llama.cpp
index 3ee95939..564304f6 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -6302,6 +6302,10 @@ static void llm_load_vocab(
                 tokenizer_pre == "bailingmoe") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 vocab.tokenizer_clean_spaces = false;
+            } else if (
+                tokenizer_pre == "seed-coder") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+                vocab.tokenizer_clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }