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 7, 12:45 UTC
[Fix] Avoid sending image input to other PP ranks (#35405)
Signed-off-by: emricksini-h <emrick.birivoutin@hcompany.ai> Co-authored-by: Roger Wang <hey@rogerw.io>
emricksini-h committed
Feb 28, 2026 at 17:14 UTC
49b9ae32e94b902b87e3d2894f5ac4a5f8dd4abb
1 file changed
+12
vllm/v1/executor/ray_utils.py
+12
@@ -104,6 +104,18 @@ try:
104
scheduler_output, intermediate_tensors
105
)
106
if self._is_intermediate_tensors(output):
107
+ if (
108
+ self.worker.model_runner.supports_mm_inputs
109
+ and get_pp_group().is_first_rank
110
+ ):
111
+ # Strip mm_features before Ray forwards it to the next PP Stage.
112
+ # PP Stage>0 only needs the intermediate tensors,
113
+ # not preprocessed multimodal data.
114
+
115
+ # scheduled_new_reqs is a required field of SchedulerOutput,
116
+ # so accessing it directly will raise AttributeError if missing.
117
+ for req in scheduler_output.scheduled_new_reqs:
118
+ req.mm_features = []
119
return scheduler_output, grammar_output, output
120
121
if isinstance(output, AsyncModelRunnerOutput):