mpt : updated convert-mpt-hf-to-gguf.py to reflect changes made to convert-gptneox-hf-to-gguf.py in pr:3252

convert-mpt-hf-to-gguf.py

@@ -19,29 +19,6 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
-# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
-
-
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a corresponding list of unicode strings.
-    The reversible bpe codes work on unicode strings.
-    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
-    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
-    This is a significant percentage of your normal, say, 32K bpe vocab.
-    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
-    And avoids mapping to whitespace/control characters the bpe code barfs on.
-    """
-    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8+n)
-            n += 1
-    return dict(zip(bs, (chr(n) for n in cs)))
-
 
 def count_model_parts(dir_model: Path) -> int:
     num_parts = 0
@@ -131,6 +108,8 @@ gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
 print("gguf: get tokenizer metadata")
 
 tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
 
 tokenizer_json_file = dir_model / 'tokenizer.json'
 if not tokenizer_json_file.is_file():
@@ -155,31 +134,15 @@ vocab_size = len(tokenizer_json["model"]["vocab"]) + 178
 tokenizer = AutoTokenizer.from_pretrained(dir_model)
 
 reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
-byte_encoder = bytes_to_unicode()
-byte_decoder = {v: k for k, v in byte_encoder.items()}
 
 for i in range(vocab_size):
-    if i in reverse_vocab:
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
-        try:
+    scores.append(0.0) # dummy
-            text = bytearray([byte_decoder[c] for c in reverse_vocab[i]])
+    toktypes.append(gguf.TokenType.NORMAL)
-        except KeyError:
-            text = bytearray()
-            for c in reverse_vocab[i]:
-                if ord(c) < 256:  # single byte character
-                    try:
-                        text.append(byte_decoder[c])
-                    except KeyError:
-                        text.extend(c.encode('utf-8'))
-                else:  # multibyte special token character
-                    text.extend(c.encode('utf-8'))
-    else:
-        print(f"Key {i} not in tokenizer vocabulary. Padding with an arbitrary token. (It's normal for MPT.)")
-        pad_token = f"[PAD{i}]".encode("utf8")
-        text = bytearray(pad_token)
-
-    tokens.append(text)
 
 gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
 
 special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
 special_vocab.add_to_gguf(gguf_writer)
@@ -239,10 +202,6 @@ for part_name in part_names:
 
         print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
 
-
-#        if new_name == "wte.weight" and data.shape[0] == 50432 and vocab_size == 50254:
-#            data = data[0:vocab_size,:]
-
         gguf_writer.add_tensor(new_name, data)
 
         # note: MPT output is tied to (same as) wte in original model;