vocab : add Carbon-3B (HybridDNATokenizer) support (#23410)

kashif · CISC · web-flow · commit 7ea23ddf7b7c · 2026-05-21T08:34:32.000+02:00
* vocab : add Carbon-3B (HybridDNATokenizer) support

Adds a new BPE pre-type LLAMA_VOCAB_PRE_TYPE_CARBON for the
HybridDNATokenizer used by HuggingFaceBio/Carbon-{500M,3B,8B}.
The base BPE is Qwen3-4B-Base's; what differs is that text inside
&lt;dna&gt;...&lt;/dna&gt; regions is chunked into fixed 6-mers (right-padded
with 'A' on the trailing partial), and any base outside ACGT maps
to &lt;oov&gt;.

* src/llama-vocab.{h,cpp}: new pre-type, dispatched from
  llm_tokenizer_bpe_session::tokenize.
* src/llama-vocab-carbon.h: pure helpers (tokenize_carbon,
  emit_dna_kmers) factored out for unit testing — no llama_vocab
  dependency, vocab access goes through a std::function.
* conversion/base.py: detect HybridDNATokenizer by class name in
  get_vocab_base_pre (chktxt collides with Qwen3 base since it
  has no &lt;dna&gt;), and pass trust_remote_code=True in get_vocab_base
  so the custom tokenizer class can load.
* tests/test-tokenizer-carbon.cpp: 12 cases covering single 6-mer,
  multi 6-mer, lowercase, invalid base -&gt; &lt;oov&gt;, partial k-mer
  right-pad, mixed text+DNA, empty &lt;dna&gt;&lt;/dna&gt;, unterminated &lt;dna&gt;,
  two regions, vocab miss.

* vocab : align Carbon-3B changes with llama.cpp conventions

* Fold tokenize_carbon + emit_dna_kmers inline into
  llm_tokenizer_bpe_session (drop src/llama-vocab-carbon.h),
  matching how every other tokenizer keeps its helpers inside
  llama-vocab.cpp.

* Replace the standalone unit test with the conventional
  test-tokenizer-0 row backed by models/ggml-vocab-carbon.gguf
  (vocab-only conversion) + .inp/.out fixtures covering single
  6-mer, multi 6-mer, lowercase, invalid base -&gt; &lt;oov&gt;, partial
  right-pad, mixed text+DNA, empty &lt;dna&gt;&lt;/dna&gt;, unterminated &lt;dna&gt;,
  two regions.

* Register "carbon" in convert_hf_to_gguf_update.py's model list
  (pointing at HuggingFaceBio/Carbon-3B) and teach both
  AutoTokenizer call sites in the updater to pass
  trust_remote_code=True for it, matching how t5 is special-cased.

* vocab : move Carbon dispatch to _set_vocab_carbon + LlamaModel branch

Refactor the conversion-side changes to follow the per-tokenizer-family
convention used by _set_vocab_qwen, _set_vocab_interns1, _set_vocab_glm,
etc. instead of conditionalising the shared get_vocab_base /
get_vocab_base_pre paths.

* conversion/base.py: add _set_vocab_carbon — self-contained, loads
  with trust_remote_code=True so HybridDNATokenizer's merged Qwen3 + DNA
  vocab is visible, writes tokenizer.ggml.pre = "carbon" directly.
* conversion/llama.py: branch in LlamaModel.set_vocab on
  tokenizer_config.json["tokenizer_class"] == "HybridDNATokenizer" and
  dispatch to _set_vocab_carbon. Same precedent as conversion/bert.py
  (tokenizer_class branch between BertTokenizer / RobertaTokenizer) and
  conversion/phi.py.
* conversion/base.py: revert the conditional in get_vocab_base and the
  class-name short-circuit in the auto-generated get_vocab_base_pre.

* tests : expand ggml-vocab-carbon.gguf fixtures with model-card examples

Add 6 cases from the Carbon-3B model card on top of the existing edge
coverage: the unterminated basic-completion prompt, the closed 33-bp
example, the metadata-conditioned prompt (with &lt;vertebrate_mammalian&gt;
and &lt;protein_coding_region&gt; which BPE-decompose since they are not in
the vocab), the documented anti-pattern of raw DNA without &lt;dna&gt; tags,
and the two likelihood-scoring examples. Brings the suite to 19 cases.

* vocab : promote HybridDNATokenizer to its own LLAMA_VOCAB_TYPE

Refactor per upstream review:

&gt; This should be its own tokenizer model, ie. carbonhybriddna instead
&gt; of gpt2 and not carbon pre-tokenizer. That way you can keep the
&gt; correct pre-tokenizer, in case that ever changes.

Previously the tokenizer was modelled as LLAMA_VOCAB_TYPE_BPE plus a
new LLAMA_VOCAB_PRE_TYPE_CARBON, which (a) put a CARBON-specific
branch inside llm_tokenizer_bpe_session::tokenize (only existing
pre-types differ in regex, not dispatch logic), and (b) conflated
"hybrid DNA tokenization" with "Qwen3 BPE pre-tokenizer".

This change moves it to its own vocab type, peer to PLAMO2, with the
GGUF model name matching the HF tokenizer class (HybridDNATokenizer):

* include/llama.h: new LLAMA_VOCAB_TYPE_HYBRIDDNA = 7.
* src/llama-vocab.cpp: new llm_tokenizer_hybriddna + session that
  owns std::unique_ptr&lt;llm_tokenizer_bpe&gt; for non-&lt;dna&gt; text and
  routes raw text through a DNA-aware splitter; wired into
  init_tokenizer, tokenize, type_name, byte_to_token, and the
  BPE-style token_to_piece case (DNA k-mers + &lt;dna&gt;/&lt;/dna&gt;/&lt;oov&gt;
  are pure ASCII, so byte-level BPE decoding handles them).
  LLAMA_VOCAB_TYPE_HYBRIDDNA gets its own branch in the vocab-type
  config block alongside SPM/WPM/UGM/RWKV, where pre_type is set
  to QWEN2 and the matching add_space_prefix / escape_whitespaces /
  clean_spaces flags are applied — mirroring qwen2's BPE path so
  byte-level BPE merging stays bit-identical to the Python
  reference for non-DNA text.
* src/llama-vocab.h: drop the short-lived LLAMA_VOCAB_PRE_TYPE_CARBON.
* conversion/base.py: _set_vocab_hybriddna writes
  tokenizer.ggml.model = "hybriddna" (no separate pre).
* conversion/llama.py: dispatch on tokenizer_class ==
  "HybridDNATokenizer" same as bert.py / phi.py do.
* models/ggml-vocab-hybriddna.gguf{,.inp,.out}: renamed fixture +
  regenerated metadata.
* convert_hf_to_gguf_update.py: drop the stale chkhsh entry and
  trust_remote_code special-case (no longer needed since dispatch
  is now class-name driven, not chkhsh).

Verified end-to-end against HuggingFaceBio/Carbon-{500M,3B,8B}:
tokenization is bit-identical to the Python HybridDNATokenizer for
all 19 test fixtures plus the model-card metadata-conditioned
prompt; greedy completion produces the same DNA continuation as
the Python reference; spec-dec with 500M as draft for 8B still
works.

* vocab : relax llm_tokenizer_bpe assert to allow HYBRIDDNA

* vocab : drop llm_tokenizer_bpe vocab-type assert

* vocab : write tokenizer.ggml.pre for HYBRIDDNA, share BPE dispatch

* vocab : assert BPE or HYBRIDDNA in llm_tokenizer_bpe

* vocab : annotate #endif with PRETOKENIZERDEBUG

* vocab : drop local hybriddna fixture (moves to ggml-org/vocabs)

* deduplicate

* simplify

* simplify

---------

Co-authored-by: Sigbjørn Skjæret &lt;sigbjorn.skjaeret@scala.com&gt;
diff --git a/conversion/base.py b/conversion/base.py
@@ -1610,6 +1610,42 @@ def _set_vocab_gpt2(self) -> None:
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
         special_vocab.add_to_gguf(self.gguf_writer)
 
+    def _set_vocab_hybriddna(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+
+        tokens: list[str] = []
+        toktypes: list[int] = []
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+        self.gguf_writer.add_tokenizer_model("hybriddna")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def _set_vocab_qwen(self):
         from .qwen import QwenModel
 
diff --git a/conversion/llama.py b/conversion/llama.py
@@ -51,6 +51,15 @@ def set_vocab(self):
         if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
             self._set_vocab_mistral()
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if (add_prefix_space := tokenizer_config_json.get("add_prefix_space")) is not None:
+                    self.gguf_writer.add_add_space_prefix(add_prefix_space)
+                if tokenizer_config_json.get("tokenizer_class") == "HybridDNATokenizer":
+                    return self._set_vocab_hybriddna()
+
         try:
             self._set_vocab_sentencepiece()
         except FileNotFoundError:
@@ -72,13 +81,6 @@ def set_vocab(self):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
         # Apply to granite small models only
         if self.hparams.get("vocab_size", 32000) == 49152:
             self.gguf_writer.add_add_bos_token(False)
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
@@ -530,6 +530,8 @@ struct llm_tokenizer_bpe : llm_tokenizer {
 struct llm_tokenizer_bpe_session {
     llm_tokenizer_bpe_session(const llama_vocab & vocab, const llm_tokenizer_bpe & tokenizer) : vocab(vocab), tokenizer(tokenizer) {}
 
+    virtual ~llm_tokenizer_bpe_session() = default;
+
     static void append(const llama_token token_id, std::vector<llama_token> & output)  {
         output.push_back(token_id);
     }
@@ -567,7 +569,7 @@ struct llm_tokenizer_bpe_session {
         }
     }
 
-    void tokenize(const std::string & text, std::vector<llama_token> & output) {
+    virtual void tokenize(const std::string & text, std::vector<llama_token> & output) {
         int final_prev_index = -1;
         const auto word_collection = unicode_regex_split(text, tokenizer.regex_exprs, tokenizer.byte_encode);
 
@@ -1579,6 +1581,95 @@ struct llm_tokenizer_plamo2_session {
     const llm_tokenizer_plamo2 & tokenizer;
 };
 
+struct llm_tokenizer_hybriddna_session : llm_tokenizer_bpe_session {
+    llm_tokenizer_hybriddna_session(const llama_vocab & vocab, const llm_tokenizer_bpe & tokenizer) : llm_tokenizer_bpe_session{vocab, tokenizer}, vocab{vocab} {}
+
+    void tokenize(const std::string & text, std::vector<llama_token> & output) override {
+        static const std::string open_tag  = "<dna>";
+        static const std::string close_tag = "</dna>";
+
+        const auto dna_begin_id = vocab.text_to_token(open_tag);
+        const auto dna_end_id   = vocab.text_to_token(close_tag);
+        const auto dna_oov_id   = vocab.text_to_token("<oov>");
+
+        // Fall back to plain BPE if the DNA pieces aren't in the vocab.
+        if (dna_begin_id == LLAMA_TOKEN_NULL || dna_end_id == LLAMA_TOKEN_NULL || dna_oov_id == LLAMA_TOKEN_NULL) {
+            llm_tokenizer_bpe_session::tokenize(text, output);
+            return;
+        }
+
+        const size_t k = 6;
+        size_t pos = 0;
+
+        while (pos < text.size()) {
+            const size_t start = text.find(open_tag, pos);
+            if (start == std::string::npos) {
+                if (pos < text.size()) {
+                    llm_tokenizer_bpe_session::tokenize(text.substr(pos), output);
+                }
+                break;
+            }
+            if (start > pos) {
+                llm_tokenizer_bpe_session::tokenize(text.substr(pos, start - pos), output);
+            }
+            output.push_back(dna_begin_id);
+
+            const size_t content_start = start + open_tag.size();
+            const size_t end           = text.find(close_tag, content_start);
+            const size_t content_end   = (end == std::string::npos) ? text.size() : end;
+
+            emit_dna_kmers(text.substr(content_start, content_end - content_start), k, dna_oov_id, output);
+
+            if (end == std::string::npos) {
+                break;
+            }
+            output.push_back(dna_end_id);
+            pos = end + close_tag.size();
+        }
+    }
+
+private:
+    void emit_dna_kmers(const std::string & raw, size_t k, llama_token oov_id, std::vector<llama_token> & output) {
+        std::string seq = raw;
+        for (char & c : seq) {
+            if (c >= 'a' && c <= 'z') {
+                c = char(c - 32);
+            }
+        }
+        auto is_valid_kmer = [](const std::string & s) {
+            for (char c : s) {
+                if (c != 'A' && c != 'C' && c != 'G' && c != 'T') {
+                    return false;
+                }
+            }
+            return true;
+        };
+
+        size_t i = 0;
+        for (; i + k <= seq.size(); i += k) {
+            const std::string kmer = seq.substr(i, k);
+            if (is_valid_kmer(kmer)) {
+                const auto tok = vocab.text_to_token(kmer);
+                output.push_back(tok != LLAMA_TOKEN_NULL ? tok : oov_id);
+            } else {
+                output.push_back(oov_id);
+            }
+        }
+        if (i < seq.size()) {
+            std::string kmer = seq.substr(i);
+            kmer.append(k - kmer.size(), 'A');
+            if (is_valid_kmer(kmer)) {
+                const auto tok = vocab.text_to_token(kmer);
+                output.push_back(tok != LLAMA_TOKEN_NULL ? tok : oov_id);
+            } else {
+                output.push_back(oov_id);
+            }
+        }
+    }
+
+    const llama_vocab & vocab;
+};
+
 //
 // impl
 //
@@ -1808,7 +1899,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             special_mask_id = 103;
 
             add_sep = true;
-        } else if (tokenizer_model == "gpt2") {
+        } else if (tokenizer_model == "gpt2" || tokenizer_model == "hybriddna") {
             type = LLAMA_VOCAB_TYPE_BPE;
 
             // read bpe merges and populate bpe ranks
@@ -3144,11 +3235,19 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
             } break;
         case LLAMA_VOCAB_TYPE_BPE:
             {
-                llm_tokenizer_bpe_session session(vocab, *static_cast<const llm_tokenizer_bpe *>(tokenizer.get()));
                 // it calls some other methods that are not exist in llm_tokenizer,
                 // here just cast it to bpe tokenizer object
+                const llm_tokenizer_bpe * tok_bpe = static_cast<const llm_tokenizer_bpe *>(tokenizer.get());
+
+                std::unique_ptr<llm_tokenizer_bpe_session> session;
+                if (vocab.get_tokenizer_model() == "hybriddna") {
+                    session = std::make_unique<llm_tokenizer_hybriddna_session>(vocab, *tok_bpe);
+                } else {
+                    session = std::make_unique<llm_tokenizer_bpe_session>(vocab, *tok_bpe);
+                }
+
                 if (add_special) {
-                    session.append_bos(output);
+                    session->append_bos(output);
                 }
                 for (const auto & fragment : fragment_buffer) {
                     if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_RAW_TEXT) {
@@ -3161,15 +3260,15 @@ std::vector<llama_token> llama_vocab::impl::tokenize(
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("TT: (%ld %ld %ld) '%s'\n", text.length(), fragment.offset, fragment.length, text.c_str());
 #endif
-                        session.tokenize(text, output);
+                        session->tokenize(text, output);
                     } else { // if (fragment.type == FRAGMENT_BUFFER_VARIANT_TYPE_TOKEN)
-                        session.append(fragment.token, output);
+                        session->append(fragment.token, output);
                     }
                 }
 
                 if (add_special) {
-                    session.append_eos(output);
-                    session.check_double_bos_eos(output);
+                    session->append_eos(output);
+                    session->check_double_bos_eos(output);
                 }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
