From c0fdfd4b92d316170830c9a056769629e06b1012 Mon Sep 17 00:00:00 2001 From: siyu Date: Wed, 25 Feb 2026 10:13:03 +0800 Subject: [PATCH] Delete mm.feature after decode phase (#17324) --- .../srt/managers/scheduler_output_processor_mixin.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/python/sglang/srt/managers/scheduler_output_processor_mixin.py b/python/sglang/srt/managers/scheduler_output_processor_mixin.py index b0883ae29..ee0409895 100644 --- a/python/sglang/srt/managers/scheduler_output_processor_mixin.py +++ b/python/sglang/srt/managers/scheduler_output_processor_mixin.py @@ -453,6 +453,13 @@ class SchedulerOutputProcessorMixin: req.check_finished(new_accepted_len) if req.finished(): + # delete feature to save memory + if req.multimodal_inputs is not None: + for mm_item in req.multimodal_inputs.mm_items: + pixel_values = mm_item.feature + if isinstance(pixel_values, torch.Tensor): + mm_item.feature = None + del pixel_values self.maybe_collect_routed_experts(req) if self.server_args.disaggregation_decode_enable_offload_kvcache: