bugfix: assertions, wrong special token list

Author: jaime-m-p

llama.cpp

@@ -4903,16 +4903,19 @@ static void llm_load_vocab(
                 return false;
             };
 
-            auto _set_token_attrib = [&vocab] (const std::string & token, llama_token_attrib attrib, bool value) {
-                llama_vocab::id id = vocab.token_to_id.at(token);
+            auto _set_tokenid_attrib = [&] (const llama_vocab::id id, llama_token_attrib attrib, bool value) {
                 uint32_t attribs = vocab.id_to_token[id].attribs;
                 attribs = value ? (attribs | attrib) : (attribs & ~attrib);
                 vocab.id_to_token[id].attribs = (llama_token_attrib) attribs;
             };
 
+            auto _set_token_attrib = [&] (const std::string & token, llama_token_attrib attrib, bool value) {
+                _set_tokenid_attrib(vocab.token_to_id.at(token), attrib, value);
+            };
+
             if (_contains_any({"phi-3", "phi3"})) {
-                for (auto token : vocab.cache_token_to_piece_special) {
-                    _set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, true);
+                for (auto id : vocab.cache_special_tokens) {
+                    _set_tokenid_attrib(id, LLAMA_TOKEN_ATTRIB_RSTRIP, true);
                 }
                 for (auto token : {"</s>"}) {
                     _set_token_attrib(token, LLAMA_TOKEN_ATTRIB_RSTRIP, true);
@@ -13312,7 +13315,8 @@ struct fragment_buffer_variant {
 static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<fragment_buffer_variant> & buffer) {
     // for each special token
     for (const llama_vocab::id special_id : vocab.cache_special_tokens) {
-        const auto & special_token = vocab.id_to_token[special_id].text;
+        const auto & data = vocab.id_to_token[special_id];
+        const auto & special_token = data.text;
 
         // for each text fragment
         std::forward_list<fragment_buffer_variant>::iterator it = buffer.begin();
@@ -13349,13 +13353,22 @@ static void tokenizer_st_partition(const llama_vocab & vocab, std::forward_list<
                     if (match > raw_text_base_offset) {
                         // left
                         const int64_t left_reminder_offset = raw_text_base_offset + 0;
-                        const int64_t left_reminder_length = match - raw_text_base_offset;
-                        buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length);
+                        int64_t left_reminder_length = match - raw_text_base_offset;
+
+                        if (data.attribs & LLAMA_TOKEN_ATTRIB_LSTRIP) {
+                            while (left_reminder_length > 0 && isspace(raw_text[left_reminder_offset + left_reminder_length - 1])) {
+                                left_reminder_length--;
+                            }
+                        }
+
+                        if (left_reminder_length > 0) {
+                            buffer.emplace_after(it, raw_text, left_reminder_offset, left_reminder_length);
+                            it++;
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         LLAMA_LOG_WARN("FL: (%ld %ld) '%s'\n", left_reminder_offset, left_reminder_length, raw_text->substr(left_reminder_offset, left_reminder_length).c_str());
 #endif
-                        it++;
                     }
 
                     // special token