llama : validate special token ids are in range when loading GGUF model (#3635)

* Add validation for special token ids to llama.cpp

Small optimization for llama_byte_to_token SPM mode

* Fix BPE newline check, only I could break something so simple

* Killll meeeeee

* Account for GGUF_KEY_KEY only setting when the key exists

* Minor code cleanups.

* Fix convert.py error msg when added tokens are out of range

* Make gguf SpecialVocab vocab size-aware

Update conversion scripts accordingly

* Avoid a string copy

Co-authored-by: Georgi Gerganov <[email protected]>

---------

Co-authored-by: Georgi Gerganov <[email protected]>

Author: KerfuffleV2

convert-baichuan-hf-to-gguf.py

@@ -230,7 +230,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model)
+special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-bloom-hf-to-gguf.py

@@ -129,7 +129,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-falcon-hf-to-gguf.py

@@ -152,7 +152,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-gptneox-hf-to-gguf.py

@@ -134,7 +134,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-llama-ggml-to-gguf.py

@@ -388,7 +388,9 @@ def handle_metadata(cfg, hp):
         cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
         cfg.vocabtype )
     # FIXME: Respect cfg.vocab_dir?
-    svocab = gguf.SpecialVocab(cfg.model_metadata_dir)
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
+        load_merges = cfg.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
     return (params, vocab, svocab)
 

convert-mpt-hf-to-gguf.py

@@ -139,7 +139,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-refact-hf-to-gguf.py

@@ -150,7 +150,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert-starcoder-hf-to-gguf.py

@@ -122,7 +122,7 @@ def parse_args() -> argparse.Namespace:
 gguf_writer.add_token_scores(scores)
 gguf_writer.add_token_types(toktypes)
 
-special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
 special_vocab.add_to_gguf(gguf_writer)
 
 # TENSORS

convert.py

@@ -369,7 +369,7 @@ def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> No
         expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids   = sorted(added_tokens.values())
         if expected_ids != actual_ids:
-            raise Exception(f"Expected added token IDs to be sequential and start at {len(added_tokens)}; got {actual_ids}")
+            raise Exception(f"Expected added token IDs to be sequential and start at {vocab_size}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_list = [text for (text, idx) in items]
@@ -1163,10 +1163,13 @@ def main(args_in: list[str] | None = None) -> None:
 
     vocab: Vocab
     if args.vocab_only:
-        assert args.outfile, "need --outfile if using --vocab-only"
+        if not args.outfile:
+            raise ValueError("need --outfile if using --vocab-only")
         # FIXME: Try to respect vocab_dir somehow?
         vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
-        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+            load_merges = args.vocabtype == 'bpe',
+            n_vocab = vocab.vocab_size)
         outfile = args.outfile
         OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
@@ -1178,7 +1181,9 @@ def main(args_in: list[str] | None = None) -> None:
         vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
         vocab = load_vocab(vocab_dir, args.vocabtype)
     # FIXME: Try to respect vocab_dir somehow?
-    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, load_merges = args.vocabtype == 'bpe')
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+        load_merges = args.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
 
     model   = model_plus.model
     model   = convert_model_names(model, params)

gguf-py/gguf/gguf.py

@@ -987,12 +987,15 @@ class SpecialVocab:
     merges: list[str] = []
     special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad')
     special_token_ids: dict[str, int] = {}
+    n_vocab: int | None = None
 
     def __init__(
         self, path: str | os.PathLike[str], load_merges: bool = False,
         special_token_types: tuple[str, ...] | None = None,
+        n_vocab: int | None = None,
     ):
         self.special_token_ids = {}
+        self.n_vocab = n_vocab
         self.load_merges = load_merges
         if special_token_types is not None:
             self.special_token_types = special_token_types
@@ -1002,6 +1005,16 @@ def _load(self, path: Path) -> None:
         if not self._try_load_from_tokenizer_json(path):
             self._try_load_from_config_json(path)
 
+    def _set_special_token(self, typ: str, tid: Any):
+        if not isinstance(tid, int) or tid < 0:
+            return
+        if self.n_vocab is None or tid < self.n_vocab:
+            self.special_token_ids[typ] = tid
+            return
+        print(f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping',
+            file = sys.stderr)
+
+
     def _try_load_from_tokenizer_json(self, path: Path) -> bool:
         tokenizer_file = path / 'tokenizer.json'
         if not tokenizer_file.is_file():
@@ -1029,10 +1042,11 @@ def _try_load_from_tokenizer_json(self, path: Path) -> bool:
                 tc_content = entry_content
             else:
                 continue
-            for maybe_token_id in (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content):
-                if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
-                    self.special_token_ids[typ] = maybe_token_id
-                break
+            # We only need the first match here.
+            maybe_token_id = next((
+                atok.get('id') for atok in added_tokens
+                if atok.get('content') == tc_content), None)
+            self._set_special_token(typ, maybe_token_id)
         return True
 
     def _try_load_from_config_json(self, path: Path) -> bool:
@@ -1042,21 +1056,21 @@ def _try_load_from_config_json(self, path: Path) -> bool:
         with open(config_file, encoding = 'utf-8') as f:
             config = json.load(f)
         for typ in self.special_token_types:
-            maybe_token_id = config.get(f'{typ}_token_id')
-            if isinstance(maybe_token_id, int) and maybe_token_id >= 0:
-                self.special_token_ids[typ] = maybe_token_id
+            self._set_special_token(typ, config.get(f'{typ}_token_id'))
         return True
 
-    def add_to_gguf(self, gw: GGUFWriter) -> None:
+    def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None:
         if len(self.merges) > 0:
-            print(f'gguf: Adding {len(self.merges)} merge(s).')
+            if not quiet:
+                print(f'gguf: Adding {len(self.merges)} merge(s).')
             gw.add_token_merges(self.merges)
         for typ, tokid in self.special_token_ids.items():
             handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None)
             if handler is None:
-                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping')
+                print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', file = sys.stderr)
                 continue
-            print(f'gguf: Setting special token type {typ} to {tokid}')
+            if not quiet:
+                print(f'gguf: Setting special token type {typ} to {tokid}')
             handler(tokid)
 
     def __repr__(self) -> str:

llama.cpp

@@ -2238,15 +2238,35 @@ static void llm_load_vocab(
     if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
         vocab.linefeed_id = llama_byte_to_token(vocab, '\n');
     } else {
-        vocab.linefeed_id = llama_tokenize_internal(vocab, "\u010A", false)[0];
+        const std::vector<int> ids = llama_tokenize_internal(vocab, "\u010A", false);
+        GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
+        vocab.linefeed_id = ids[0];
     }
 
     // special tokens
-    GGUF_GET_KEY(ctx, vocab.special_bos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_BOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_eos_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_EOS_ID));
-    GGUF_GET_KEY(ctx, vocab.special_unk_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_UNK_ID));
-    GGUF_GET_KEY(ctx, vocab.special_sep_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_SEP_ID));
-    GGUF_GET_KEY(ctx, vocab.special_pad_id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_TOKENIZER_PAD_ID));
+    {
+        const std::vector<std::pair<enum llm_kv, int32_t &>> special_token_types = {
+            { LLM_KV_TOKENIZER_BOS_ID, vocab.special_bos_id },
+            { LLM_KV_TOKENIZER_EOS_ID, vocab.special_eos_id },
+            { LLM_KV_TOKENIZER_UNK_ID, vocab.special_unk_id },
+            { LLM_KV_TOKENIZER_SEP_ID, vocab.special_sep_id },
+            { LLM_KV_TOKENIZER_PAD_ID, vocab.special_pad_id },
+        };
+        for (const auto & it : special_token_types) {
+            const std::string & key = kv(std::get<0>(it));
+            int32_t & id = std::get<1>(it), old_id = id;
+
+            GGUF_GET_KEY(ctx, id, gguf_get_val_u32, GGUF_TYPE_UINT32, false, key);
+            // Must be >= -1 and < vocab size. Since the key is unsigned, -1
+            // can only come from the default value, so there's no point in
+            // validating that.
+            if (size_t(id + 1) > vocab.id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: bad special token: '%s' = %d, using default id %d\n",
+                    __func__, key.c_str(), id, old_id);
+                id = old_id;
+            }
+        }
+    }
 
     // build special tokens cache
     {
@@ -6103,11 +6123,10 @@ static uint8_t llama_token_to_byte(const llama_vocab& vocab, llama_token id) {
 }
 
 static llama_token llama_byte_to_token(const llama_vocab & vocab, uint8_t ch) {
+    static const char * hex = "0123456789ABCDEF";
     switch (llama_vocab_get_type(vocab)) {
     case LLAMA_VOCAB_TYPE_SPM: {
-        char buf[7];
-        int result = snprintf(buf, sizeof(buf), "<0x%02X>", ch);
-        GGML_ASSERT(0 <= result && result < 7);
+        const char buf[7] = { '<', '0', 'x', hex[ch >> 4], hex[ch & 15], '>', 0 };
         return vocab.token_to_id.at(buf);
     }
     case LLAMA_VOCAB_TYPE_BPE: {