@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 68ff30d40e
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

[Bugfix] Fixes MiniCPM-O resampler device placement to avoid tensor device mismatch (#42332)

Signed-off-by: j9smith <j.smith9103@outlook.com>

Joel Smith committed Jun 17, 2026 at 09:35 UTC 68ff30d40e9f0ef1eaff33f99a155f719ccb7bb9
1 file changed +3 -1
vllm/model_executor/models/minicpmo.py
+3 -1
@@ -719,7 +719,9 @@ class MiniCPMOBaseModel:
719
720 def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
721 loader = AutoWeightsLoader(self, skip_prefixes=["tts"])
722 - return loader.load_weights(weights)
722 + loaded = loader.load_weights(weights)
723 + self._ensure_resampler_device()
724 + return loaded
725
726 def subsequent_chunk_mask(
727 self,