[MoE Refactor] Remove MoE legacy code (#51078)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-08-05 14:35:36 -07:00
committed by GitHub
parent 4282fe52d5
commit 373fe8b83e
22 changed files with 3 additions and 295 deletions
@@ -35,7 +35,6 @@ from vllm.distributed.eplb.eplb_communicator import create_eplb_communicator
from vllm.distributed.parallel_state import (
_replace_active_groups,
get_eplb_group,
prepare_communication_buffer_for_model,
)
from vllm.distributed.stateless_coordinator import StatelessGroupCoordinator
from vllm.logger import init_logger
@@ -485,7 +484,6 @@ class ElasticEPScalingExecutor:
for module in moe_modules:
if getattr(module._quant_method, "wraps_legacy_quant_method", False):
module._replace_quant_method(module._quant_method.old_quant_method)
prepare_communication_buffer_for_model(self.worker.model_runner.model)
eplb_model_state.expert_buffer = [
torch.empty_like(w) for w in model.expert_weights[0]
@@ -652,10 +650,6 @@ class ElasticEPScalingExecutor:
old_num_physical_experts,
)
def prepare_new_worker(self) -> None:
with set_current_vllm_config(self.worker.vllm_config):
prepare_communication_buffer_for_model(self.worker.model_runner.get_model())
def warmup_local_kernels(self) -> None:
with set_current_vllm_config(self.worker.vllm_config):
kernel_warmup(self.worker, process_local_only=True)