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 <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

1 files changed, 28 insertions, 9 deletions

diff --git a/llama.cpp b/llama.cpp
index 36534933..8d52eaf6 100644
--- a/llama.cpp
+++ b/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: {