From 320c29c2dbe3c8df56374a9ec19a7fe5c124d4f8 Mon Sep 17 00:00:00 2001 From: Piotr Wilkin Date: Tue, 7 Apr 2026 00:54:00 +0200 Subject: [PATCH 1/2] YATF (Yet Another Tokenizer Fix) for Gemma 4. With tests! --- convert_hf_to_gguf_update.py | 1 + models/ggml-vocab-gemma-4.gguf | Bin 0 -> 15776467 bytes models/ggml-vocab-gemma-4.gguf.inp | 111 +++++++++++++++++++++++++++++ models/ggml-vocab-gemma-4.gguf.out | 46 ++++++++++++ src/llama-vocab.cpp | 13 +++- tests/CMakeLists.txt | 1 + 6 files changed, 170 insertions(+), 2 deletions(-) create mode 100644 models/ggml-vocab-gemma-4.gguf create mode 100644 models/ggml-vocab-gemma-4.gguf.inp create mode 100644 models/ggml-vocab-gemma-4.gguf.out diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index 086f1c22863..f1d70d62e73 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -114,6 +114,7 @@ class TOKENIZER_TYPE(IntEnum): {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, + {"name": "gemma-4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/google/gemma-4-E2B-it", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, {"name": "jais-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inceptionai/Jais-2-8B-Chat", }, {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", }, diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index de9a9466bc7..e9e276ab999 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -658,9 +658,18 @@ struct llm_tokenizer_bpe_session { const auto token = vocab.text_to_token(str); if (token == LLAMA_TOKEN_NULL) { + static const char * hex = "0123456789ABCDEF"; for (auto j = str.begin(); j != str.end(); ++j) { - std::string byte_str(1, *j); - auto token_multibyte = vocab.text_to_token(byte_str); + llama_token token_multibyte = LLAMA_TOKEN_NULL; + if (tokenizer.byte_encode) { + std::string byte_str(1, *j); + token_multibyte = vocab.text_to_token(byte_str); + } else { + // For non-byte-encoded BPE (e.g. gemma-4), byte tokens use <0xXX> format + const uint8_t ch = (uint8_t)*j; + const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 }; + token_multibyte = vocab.text_to_token(buf); + } if (token_multibyte != LLAMA_TOKEN_NULL) { output.push_back(token_multibyte); } diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 5e87c8b34e1..cd4bc5ef1d3 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -124,6 +124,7 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r ARGS ${PROJE llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-coder.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-llm.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-falcon.gguf) +llama_test(test-tokenizer-0 NAME test-tokenizer-0-gemma-4 ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gemma-4.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2 ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-2.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf) llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-spm.gguf) From 0e98596dec124c6968132ef042c21ccdb20d1304 Mon Sep 17 00:00:00 2001 From: Piotr Wilkin Date: Tue, 7 Apr 2026 00:58:08 +0200 Subject: [PATCH 2/2] Remove unnecessary hash from update script. --- convert_hf_to_gguf_update.py | 1 - 1 file changed, 1 deletion(-) diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index f1d70d62e73..086f1c22863 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -114,7 +114,6 @@ class TOKENIZER_TYPE(IntEnum): {"name": "viking", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LumiOpen/Viking-7B", }, # Also used for Viking 13B and 33B {"name": "gemma", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", }, {"name": "gemma-2", "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", }, - {"name": "gemma-4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/google/gemma-4-E2B-it", }, {"name": "jais", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", }, {"name": "jais-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inceptionai/Jais-2-8B-Chat", }, {"name": "t5", "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },