EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Dynamic Expert Load Balance with Zero-like-overhead (#2956)

Browse Source 
### Motivation
Currently dynamically experts balancing would stop-the-world.
Asynchronously expert load balancing would be better without flowing
problems:

Host-bound latency:
There are many cpu operations during EPLB such as
eplb-algorithm、creating p2p ops、and log2phy expert converting would
spend long cpu time, as ~1s.
Communication latency: The transfer time would cost much in the
situation without nvlink. As the weight of an expert maybe transfer to
multiple new positions, thus N times send/recv for one expert, with
result long latency. We had tested that batch_isend_irecv cost more
100ms for 16 experts weight transmission in A2 server of ascend.

SwiftBalancer would not stop-the-world anymore, in out test on NPU 1~2ms
cost for each layer while benefit 5ms-8ms decode latency with ep_size =
64.
The following updates have been made:
1、expert distribution recording with lower cost.
2、async cpu computing for eplb algo and other python operator.
3、new eplb algo with less expert rebalancing while almost the same
effect.
### Proposed Change
We will gradually migrate the EPLB logic to the VLLM community and
implement a generalized design. Relevant RFC:
https://github.com/vllm-project/vllm/issues/22246
The overall workflow involves:
<img width="801" height="302"
alt="474430541-23b06f58-23bc-44a3-a1be-00f268aeb15c"
src="https://github.com/user-attachments/assets/1d73a459-1b23-4b0a-812a-bf0a75debfed"
/>
1. Record experts distribution during forward. We using expert_token_num
after disptach instead of topk_ids, thus we got much smaller tensor
shape to reduce cost of hbm recording and add-operator.
2. Do all-gather for experts distribution. Using all-gather instead of
all-reduce as less traffic volume.
3. Wake up eplb worker process with experts distribution when
num_iterations comes. Run eplb algorithm in eplb worker.
4. Generate p2p send/recv ops and other operator such as log2phy would
cost long cpu time.
5. Lanch ibatch_send_recv in async_stream before forward.
6. After forward, wait for the ibatch_send_recv finish, then do uapte
expert map and expert weights.
### Co-author
Co-authored-by: raindaywhu raindaywhu@raindaywhu@ 163.con
Co-authored-by: njuyuan yuanjl19@smail.nju.edu.cn
Co-authored-by: qmkakaxi wjh1594260677@qq.com
Co-authored-by: Skywalker-EP 173723846@qq.com


- vLLM version: v0.10.2
- vLLM main:
567939953b

---------

Signed-off-by: offline0806 <z00858301@china.huawei.com>
Co-authored-by: offline0806 <z00858301@china.huawei.com>
This commit is contained in:
offline893
2025-09-17 10:36:43 +08:00
committed by GitHub
parent ae758dda05
commit 76844eec78
30 changed files with 2891 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
vllm_ascend/torchair/ops/torchair_fused_moe.py
View File

@@ -41,6 +41,8 @@ from vllm.model_executor.layers.quantization.base_config import \
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import FusedMoEState
from vllm_ascend.distributed.parallel_state import get_mc2_group
from vllm_ascend.eplb.core.eplb_utils import (determine_default_expert_map,
determine_default_log2phy_map)
from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
from vllm_ascend.ops.sequence_parallel import MetadataForPadding
from vllm_ascend.quantization.quant_config import AscendFusedMoEMethod
@@ -1011,25 +1013,40 @@ class TorchairAscendFusedMoE(FusedMoE):
self.moe_parallel_config.ep_size, is_deepseek_v3_r1)
ascend_config = get_ascend_config()
expert_map_path = ascend_config.expert_map_path
if expert_map_path and os.path.exists(expert_map_path):
# moe expert load balance
expert_load_balancer = ExpertLoadBalancer(expert_map_path,
self.global_num_experts)
self.local_num_experts, self.expert_map = \
expert_load_balancer.get_rank_placement_map(
self.moe_instance_id,
get_ep_group().rank_in_group)
self.log2phy = expert_load_balancer.get_rank_log2phy_map(
self.moe_instance_id,
get_ep_group().rank_in_group)
self.global_redundant_expert_num = \
expert_load_balancer.get_global_redundant_expert_num()
self.dynamic_eplb = ascend_config.dynamic_eplb
self.expert_map_path = ascend_config.expert_map_path
self.global_redundant_expert_num = ascend_config.init_redundancy_expert
self.global_num_experts = num_experts + self.global_redundant_expert_num
# static eplb initializing with expert_map_path
if self.expert_map_path and os.path.exists(
self.expert_map_path) and os.access(self.expert_map_path,
os.R_OK):
self.expert_load_balancer = ExpertLoadBalancer(
self.expert_map_path, self.global_num_experts)
self.local_num_experts, self.expert_map = (
self.expert_load_balancer.get_rank_placement_map(
self.moe_instance_id, self.ep_rank))
self.log2phy = self.expert_load_balancer.get_rank_log2phy_map(
self.moe_instance_id, self.ep_rank).npu()
self.global_redundant_expert_num = (
self.expert_load_balancer.get_global_redundant_expert_num())
else:
# Create a tensor of size num_experts filled with -1
# init moe.
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size,
get_ep_group().rank_in_group, self.global_num_experts)
self.ep_size, self.ep_rank, self.global_num_experts)
# dynamic eplb initializing with not expert_map_path
if self.dynamic_eplb:
self.global_redundant_expert_num = ascend_config.init_redundancy_expert
self.local_num_experts, self.expert_map = determine_default_expert_map(
self.global_num_experts, self.ep_size, self.ep_rank,
self.global_redundant_expert_num)
self.log2phy = determine_default_log2phy_map(
self.global_num_experts, self.ep_size, self.ep_rank,
self.global_redundant_expert_num)
local_num_experts = (torch.sum(self.expert_map != -1)
if self.expert_map is not None else num_experts)
if self.dynamic_eplb:
self.moe_load = torch.zeros(local_num_experts, dtype=torch.int64)
self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
self.enable_multistream_moe = \
@@ -1064,8 +1081,11 @@ class TorchairAscendFusedMoE(FusedMoE):
assert self.quant_method is not None
local_num_experts = torch.sum(self.expert_map != -1) \
if self.expert_map is not None else num_experts
self.moe_load = None
local_num_experts = (torch.sum(self.expert_map != -1)
if self.expert_map is not None else num_experts)
if self.dynamic_eplb:
self.moe_load = torch.zeros(local_num_experts, dtype=torch.int64)
moe_quant_params = {
"num_experts": local_num_experts,
@@ -1244,6 +1264,11 @@ class TorchairAscendFusedMoE(FusedMoE):
if isinstance(e_hidden_states, tuple):
e_hidden_states, shared_hidden_states = e_hidden_states
if self.dynamic_eplb and isinstance(
e_hidden_states, tuple) and len(e_hidden_states) == 3:
self.moe_load += e_hidden_states[2] if e_hidden_states[1] == 0 else \
torch.cat(e_hidden_states[2][:1], e_hidden_states[2][1:] - e_hidden_states[2][:-1])
if (fused_moe_state not in [
FusedMoEState.AllGather, FusedMoEState.AllGatherEP,
FusedMoEState.NaiveMulticast
@@ -1288,6 +1313,19 @@ class TorchairAscendFusedMoE(FusedMoE):
else:
return final_hidden_states
def update_expert_map(self, new_expert_map):
self.expert_map = new_expert_map
def get_map(self):
return self.expert_map
def get_log2phy_map(self):
return self.logical_to_physical_map
def clear_moe_load(self):
if self.moe_load is not None:
self.moe_load.zero_()
# ----------------------------------------- TBO-related --------------------------------------------
def _forward_ms_fused_moe_comp(