Delete mm.feature after decode phase (#17324)
This commit is contained in:
@@ -453,6 +453,13 @@ class SchedulerOutputProcessorMixin:
|
|||||||
req.check_finished(new_accepted_len)
|
req.check_finished(new_accepted_len)
|
||||||
|
|
||||||
if req.finished():
|
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)
|
self.maybe_collect_routed_experts(req)
|
||||||
|
|
||||||
if self.server_args.disaggregation_decode_enable_offload_kvcache:
|
if self.server_args.disaggregation_decode_enable_offload_kvcache:
|
||||||
|
|||||||
Reference in New Issue
Block a user