A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Perf] Speed up function `_convert_tokens_to_string_with_added_encoders` by 13.7x (#20413)
Signed-off-by: Saurabh Misra <misra.saurabh1@gmail.com> Signed-off-by: Aseem Saxena <aseem.bits@gmail.com> Co-authored-by: codeflash-ai[bot] <148906541+codeflash-ai[bot]@users.noreply.github.com> Co-authored-by: Aseem Saxena <aseem.bits@gmail.com>
Saurabh Misra committed
Aug 20, 2025 at 13:17 UTC
bf7c99dfc40bff6844b2ae57554516922eb93b71
1 file changed
+15
-10
vllm/transformers_utils/detokenizer_utils.py
+15
-10
@@ -23,27 +23,32 @@ def _convert_tokens_to_string_with_added_encoders(
23
# NOTE(woosuk): The following code is slow because it runs a for loop over
24
# the output_tokens. In Python, running a for loop over a list can be slow
25
# even when the loop body is very simple.
26
+ # Performance improvements: avoid repeated attribute and function lookups;
27
+ # localize frequently used objects;
28
+
29
sub_texts: list[str] = []
30
current_sub_text: list[str] = []
28
- all_special_tokens = set(tokenizer.all_special_tokens)
31
+ convert_tokens_to_string = tokenizer.convert_tokens_to_string
32
+ added_vocab_set = set(tokenizer.get_added_vocab())
33
+ all_special_tokens = set(
34
+ tokenizer.all_special_tokens) if skip_special_tokens else ()
35
+
36
for token in output_tokens:
30
- if skip_special_tokens and token in all_special_tokens:
37
+ # Use precomputed set for skip-special check
38
+ if token in all_special_tokens:
39
continue
32
- if token in tokenizer.get_added_vocab():
40
+ if token in added_vocab_set:
41
if current_sub_text:
34
- sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
35
- sub_texts.append(sub_text)
36
- current_sub_text = []
42
+ sub_texts.append(convert_tokens_to_string(current_sub_text))
43
+ current_sub_text.clear()
44
sub_texts.append(token)
45
else:
46
current_sub_text.append(token)
47
if current_sub_text:
41
- sub_text = tokenizer.convert_tokens_to_string(current_sub_text)
42
- sub_texts.append(sub_text)
48
+ sub_texts.append(convert_tokens_to_string(current_sub_text))
49
if spaces_between_special_tokens:
50
return " ".join(sub_texts)
45
- else:
46
- return "".join(sub_texts)
51
+ return "".join(sub_texts)
52
53
54
# 5 is an arbitrary value that should work for all