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

eplb redundant expert bugfix (#4291)

Browse Source 
### What this PR does / why we need it?
Redundant experts bugfix
### Does this PR introduce _any_ user-facing change?
After configuring the path for experts_map, users do not need to
configure iinit_redundancy_expert.
### How was this patch tested?
The accuracy of EPLB was tested with and without the use of redundant
experts.


- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
LI SHENGYONG
2025-11-21 14:24:35 +08:00
committed by GitHub
parent 5a4e8cdeba
commit 019c7ded91
10 changed files with 63 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
vllm_ascend/ops/fused_moe/fused_moe.py
View File

@@ -32,8 +32,7 @@ from vllm.model_executor.layers.fused_moe.layer import (
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.ascend_forward_context import MoECommType
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.eplb.core.eplb_utils import determine_default_log2phy_map
from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
from vllm_ascend.ops.fused_moe.experts_selector import select_experts
from vllm_ascend.ops.fused_moe.moe_comm_method import setup_moe_comm_method
@@ -183,10 +182,8 @@ class AscendFusedMoE(FusedMoE):
AscendFusedMoE.moe_counter += 1
self.moe_instance_id = AscendFusedMoE.moe_counter
self.global_num_experts = num_experts
self.expert_map = None
self.log2phy = None
self.global_redundant_expert_num = 0
if self.quant_config is None:
self.quant_method = AscendUnquantizedFusedMoEMethod(
@@ -210,15 +207,24 @@ class AscendFusedMoE(FusedMoE):
vllm_config = get_current_vllm_config()
self.e_score_correction_bias.data = self.e_score_correction_bias.data.to(
dtype=vllm_config.model_config.dtype)
# init moe.
if vllm_version_is("0.11.0"):
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size, self.ep_rank, self.global_num_experts)
else:
self.local_num_experts, self.expert_map, _ = determine_expert_map(
self.ep_size, self.ep_rank, self.global_num_experts)
# 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.expert_map_path, num_experts)
self.expert_load_balancer.check_expert_map_tensor()
self.global_redundant_expert_num = (
self.expert_load_balancer.get_global_redundant_expert_num())
self.global_num_experts = num_experts + self.global_redundant_expert_num
try:
self.local_num_experts, self.expert_map = (
self.expert_load_balancer.get_rank_placement_map(
@@ -228,45 +234,21 @@ class AscendFusedMoE(FusedMoE):
except Exception as e:
logger.warning(
f"Init expert map of mtp/eagle when using sample.{e}")
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).npu()
if self.expert_map is not None and isinstance(
self.expert_map, torch.Tensor):
logger.info_once(
"[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
" number of experts: %s/%s. Experts local to global index map:"
" %s.", self.ep_rank, self.ep_size, self.local_num_experts,
self.global_num_experts,
get_compressed_expert_map(self.expert_map))
self.global_num_experts, self.ep_size, self.ep_rank).npu()
else:
# init moe.
if vllm_version_is("0.11.0"):
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size, self.ep_rank, self.global_num_experts)
else:
self.local_num_experts, self.expert_map, _ = determine_expert_map(
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).npu()
if self.expert_map is not None and isinstance(
self.expert_map, torch.Tensor):
logger.info_once(
"[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
" number of experts: %s/%s. Experts local to global index map:"
" %s.", self.ep_rank, self.ep_size, self.local_num_experts,
self.global_num_experts,
get_compressed_expert_map(self.expert_map))
self.global_num_experts, self.ep_size, self.ep_rank).npu()
if self.expert_map is not None and isinstance(self.expert_map,
torch.Tensor):
logger.info_once(
"[EP Rank %s/%s] Expert parallelism is enabled. Local/global"
" number of experts: %s/%s. Experts local to global index map:"
" %s.", self.ep_rank, self.ep_size, self.local_num_experts,
self.global_num_experts,
get_compressed_expert_map(self.expert_map))
local_num_experts = (torch.sum(
self.expert_map != -1) if self.expert_map is not None else
self.global_num_experts)