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, 09:45 UTC
Revert "[Bugfix] Fix MRoPE Errors in the Qwen-VL Model When Processing Pure Text (#18407)" (#18456)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
May 21, 2025 at 13:39 UTC
ad0012a0ac507973656cfcf5750d603af4a5fdcc
1 file changed
+8
vllm/worker/model_runner.py
+8
@@ -729,6 +729,8 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
729
mm_kwargs, placeholder_maps = MultiModalPlaceholderMap.from_seq_group(
730
seq_group_metadata,
731
range(positions[0], positions[0] + len(positions)))
732
+ if not mm_kwargs:
733
+ return
734
735
inter_data.multi_modal_kwargs = mm_kwargs
736
inter_data.multi_modal_placeholder_maps = placeholder_maps
@@ -739,6 +741,12 @@ class ModelInputForGPUBuilder(ModelRunnerInputBuilderBase[ModelInputForGPU]):
741
video_grid_thw = mm_kwargs.get("video_grid_thw", None)
742
audio_feature_lengths = mm_kwargs.get("audio_feature_lengths",
743
None)
744
+ assert (
745
+ image_grid_thw is not None or video_grid_thw is not None
746
+ or audio_feature_lengths is not None), (
747
+ "mrope embedding type requires multi-modal input mapper "
748
+ "returns 'image_grid_thw' or 'video_grid_thw' or "
749
+ "'audio_feature_lengths'.")
750
751
second_per_grid_ts = mm_kwargs.get("second_per_grid_ts", None)
752
use_audio_in_video = mm_kwargs.get("use_audio_in_video", False)