vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Skip literal UNUSED token checks

Browse source
This commit is contained in:
jaime-m-p 2024-07-10 00:46:19 +02:00
parent c184db74b3
commit 3eb1900e5c
1 changed files with 14 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

18
tests/test-tokenizer-random.py
View file

@ -533,10 +533,20 @@ def compare_vocabs(tokenizer1: TokenizerGroundtruth, tokenizer2: TokenizerLlamaC
if vocab1 != vocab2: if vocab1 != vocab2:
num_errors = 0 num_errors = 0
for i in range(max(len(vocab1), len(vocab2))): for i in range(max(len(vocab1), len(vocab2))):
text1 = vocab1[i] if i < len(vocab1) else "" text1 = vocab1[i] if i < len(vocab1) else None
text2 = vocab2[i] if i < len(vocab2) else "" text2 = vocab2[i] if i < len(vocab2) else None
is_unused = text1.startswith("[UNUSED_TOKEN_") # AutoTokenizer adds more unused tokens than SentencePiece ? if True: #WIP: SentencePiece adds more unused tokens than AutoTokenizer ?
if text1 != text2 and is_unused and text2: if text1 is None:
if not text2 or text2.startswith('[PAD'): # is unused ? #TODO: use toktypes
text2 = None
else:
#TODO: is "UNUSED_TOKEN_" valid for all models ?
text1 = text1.replace("[UNUSED_TOKEN_", "[PAD")
#if text1 is None or text1.startswith("[UNUSED_TOKEN_"): # is unused ?
# text1 = ""
#if text2 is None or text2.startswith('[PAD'): # is unused ?
# text2 = ""
if text1 != text2:
num_errors += 1 num_errors += 1
if num_errors < MAX_PRINT_ERRORS: if num_errors < MAX_PRINT_ERRORS:
logger.error(f" {detokenize=} id={i} expected={repr(text1)} result={repr(text2)}") logger.error(f" {detokenize=} id={i} expected={repr(text1)} result={repr(text2)}")