Tokenizer fixes ggml-org#8379

by jaime-m-p

Author: Nexesenex

src/llama.cpp

@@ -21102,14 +21102,106 @@ int32_t llama_tokenize(
     return llama_tokenize_impl(model->vocab, text, text_len, tokens, n_tokens_max, add_special, parse_special);
 }
 
-int32_t llama_token_to_piece(
-    const struct llama_model * model,
-                 llama_token   token,
-                        char * buf,
-                     int32_t   length,
-                     int32_t   lstrip,
-                        bool   special) {
-    return llama_token_to_piece_impl(model->vocab, token, buf, length, lstrip, special);
+// errors: 'c': copy, 'i': ignore, 'r': replace 0xFFFD, 'v': verbose
+static std::string llama_decode_text(const std::string & text, const char errors = 'v') {
+    std::string decoded_text;
+
+    const auto cpts = unicode_cpts_from_utf8(text);
+    for (const auto cpt : cpts) {
+        const auto utf8 = unicode_cpt_to_utf8(cpt);
+        try {
+            decoded_text += unicode_utf8_to_byte(utf8);
+        } catch (const std::out_of_range & /*e*/) {
+            switch (errors) {
+                case 'c':
+                    decoded_text += utf8;  // copy original
+                    break;
+                case 'r':
+                    decoded_text += "\xEF\xBF\xBD";  // 0xFFFD REPLACEMENT CHARACTER
+                    break;
+                case 'v':
+                    decoded_text += format("[UNK_BYTE_0x%02X]", cpt);
+                    break;
+                case 'i':
+                default:
+                    // ignore
+                    break;
+            }
+        }
+    }
+
+    return decoded_text;
+}
+
+// does not write null-terminator to buf
+int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, int32_t lstrip, bool special) {
+    // ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843
+    static const int attr_special = LLAMA_TOKEN_ATTR_UNKNOWN | LLAMA_TOKEN_ATTR_CONTROL;
+    const llama_token_attr attr = llama_token_get_attr(model, token);
+    if (!special && (attr & attr_special)) {
+        return 0;
+    }
+
+    // copy piece chars to output text buffer
+    // skip up to 'lstrip' leading spaces before copying
+    auto _try_copy = [=] (const char * token, size_t size) -> int32_t {
+        for (int32_t i = 0; i < lstrip && size && *token == ' '; ++i) {
+            token++;
+            size--;
+        }
+        if (length < (int32_t)size) {
+            return (int32_t) -size;
+        }
+        memcpy(buf, token, size);
+        return (int32_t) size;
+    };
+
+    // if we have a cache - use it
+    {
+        const auto & cache = model->vocab.cache_token_to_piece;
+
+        if (!cache.empty()) {
+            const auto & result = cache.at(token);
+            return _try_copy(result.data(), result.size());
+        }
+    }
+
+    if (0 <= token && token < llama_n_vocab(model)) {
+        const std::string & token_text = model->vocab.id_to_token[token].text;
+        switch (llama_vocab_get_type(model->vocab)) {
+            case LLAMA_VOCAB_TYPE_WPM:
+            case LLAMA_VOCAB_TYPE_SPM:
+            case LLAMA_VOCAB_TYPE_UGM: {
+                // NOTE: we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) {
+                    return _try_copy(token_text.data(), token_text.size());
+                } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
+                    std::string result = token_text;
+                    llama_unescape_whitespace(result);
+                    return _try_copy(result.data(), result.size());
+                } else if (attr & LLAMA_TOKEN_ATTR_BYTE) {
+                    char byte = (char) llama_token_to_byte(model->vocab, token);
+                    return _try_copy((char*) &byte, 1);
+                }
+                break;
+            }
+            case LLAMA_VOCAB_TYPE_BPE: {
+                // NOTE: we accept all unsupported token types,
+                // suppressing them like CONTROL tokens.
+                if (attr & (attr_special | LLAMA_TOKEN_ATTR_USER_DEFINED)) {
+                    return _try_copy(token_text.data(), token_text.size());
+                } else if (attr & LLAMA_TOKEN_ATTR_NORMAL) {
+                    std::string result = llama_decode_text(token_text, 'c');  // copy on error  //TODO: use a tokenizer variable
+                    return _try_copy(result.data(), result.size());
+                }
+                break;
+            }
+            default:
+                GGML_ASSERT(false);
+        }
+    }
+    return 0;
 }
 
 int32_t llama_detokenize(