[MoE Refactor] Make SharedExperts class for use with DefaultMoERunner (#35153)
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
@@ -410,8 +410,7 @@ class ElasticEPScalingExecutor:
|
||||
# for the new EP size by resetting quant_method to base
|
||||
for module in moe_modules:
|
||||
if hasattr(module.quant_method, "old_quant_method"):
|
||||
module.quant_method = module.quant_method.old_quant_method
|
||||
module.runner = module._init_runner()
|
||||
module._replace_quant_method(module.quant_method.old_quant_method)
|
||||
prepare_communication_buffer_for_model(self.worker.model_runner.model)
|
||||
|
||||
eplb_model_state.communicator = create_eplb_communicator(
|
||||
|
||||
Reference in New Issue
Block a user