[MM] Cache common token sequences (#53560)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-08-24 07:48:10 -07:00
committed by GitHub
parent 4ca856b0b5
commit ecfa7bb373
14 changed files with 40 additions and 23 deletions
@@ -449,10 +449,12 @@ class MossTranscribeDiarizeDummyInputsBuilder(
dummy_mm_items = self.info.parse_mm_data(dummy_mm_data)
num_audios = mm_counts.get("audio", 0)
tokenizer = self.info.get_tokenizer()
prompt = tokenizer.encode(
prompt = cached_encode(
tokenizer,
AUDIO_PLACEHOLDER * num_audios,
add_special_tokens=False,
) or tokenizer.encode(
) or cached_encode(
tokenizer,
"\n",
add_special_tokens=False,
)