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 15, 03:30 UTC
[Bugfix] Keye-VL compatibility with `tok_kwargs` (#20058) (#20353)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Jul 2, 2025 at 14:46 UTC
ba51aea65e0d2a7afca3f25caba01500fa84d655
1 file changed
+3
-1
vllm/model_executor/models/keye.py
+3
-1
@@ -1268,11 +1268,13 @@ class KeyeMultiModalProcessor(BaseMultiModalProcessor[KeyeProcessingInfo]):
1268
prompt: str,
1269
mm_data: Mapping[str, object],
1270
mm_kwargs: Mapping[str, object],
1271
+ tok_kwargs: Mapping[str, object],
1272
) -> BatchFeature:
1273
+ mm_kwargs = self.info._get_image_processor_kwargs(**mm_kwargs)
1274
return self.info.ctx.call_hf_processor(
1275
self.info.get_hf_processor(**mm_kwargs),
1276
dict(text=prompt, **mm_data),
1275
- self.info._get_image_processor_kwargs(**mm_kwargs),
1277
+ dict(**mm_kwargs, **tok_kwargs),
1278
)
1279
1280
def _get_prompt_updates(