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
[Bugfix][Multi Modal] Fix incorrect Molmo token processing (#26873)
Signed-off-by: sanghol <sanghol@allenai.org>
sangho.lee committed
Oct 15, 2025 at 02:13 UTC
8865da157bea1b25afc4ff28031858d557363290
1 file changed
+5
-2
vllm/model_executor/models/molmo.py
+5
-2
@@ -1264,13 +1264,16 @@ class MolmoMultiModalProcessor(BaseMultiModalProcessor[MolmoProcessingInfo]):
1264
) -> list[int]:
1265
processor = self.info.get_hf_processor()
1266
1267
- # Apply the chat template to the tokens
1267
+ # The chat template is already applied to the prompt tokens
1268
+ # Use message_format="none" to avoid applying it again
1269
+ # Prepend an empty space if `always_start_with_space` is True
1270
tokens = processor.processor.get_tokens_input( # type: ignore
1271
self.info.get_tokenizer().decode(prompt_tokens),
1270
- message_format=processor.message_format,
1272
+ message_format="none",
1273
always_start_with_space=processor.always_start_with_space,
1274
)
1275
1276
+ # Prepend a BOS token id to the tokens
1277
processed_data = self.info.ctx.call_hf_processor(
1278
processor, # type: ignore
1279
dict(tokens=tokens),