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

Add static EPLB (#1116)

Browse Source 
### What this PR does / why we need it?
   Add EPLB expert map import capabilities
### Does this PR introduce _any_ user-facing change?
When importing the EPLB expert map you need import expert map file by
vllm args additional_config
### How was this patch tested?
1.You need to collect expert hotness and generate an expert placement
file based on the hotness and the EPLB algorithm, or you can directly
use an existing expert placement table.
2.When launching vLLM, enable EC2 and pass the configuration via the
command-line argument:
      --additional-config '{"expert_map_path": "/xxx/xxx/xx.json"}
Co-authored-by: songshanhu07 <1763685535@qq.com>

---------

Signed-off-by: songshanhu07 <1763685535@qq.com>
Signed-off-by: Yuxiao-Xu <664988918@qq.com>
Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
Co-authored-by: songshanhu07 <1763685535@qq.com>
Co-authored-by: Xu Yuxiao <xuyuxiao2@huawei.com>
Co-authored-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
Yuxiao-Xu
2025-06-09 19:28:11 +08:00
committed by GitHub
parent cb341c7bcd
commit 6b853f15fe
6 changed files with 179 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
vllm_ascend/ops/fused_moe.py
View File

@@ -15,6 +15,7 @@
# This file is a part of the vllm-ascend project.
# Adapted from vllm/tests/kernels/test_moe.py
import os
from typing import Callable, List, Optional
import torch
@@ -34,6 +35,7 @@ from vllm.model_executor.layers.quantization.base_config import \
import vllm_ascend.envs as envs_ascend
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.distributed.parallel_state import get_ep_group, get_etp_group
from vllm_ascend.ops.expert_load_balancer import ExpertLoadBalancer
VLLM_ENABLE_MC2: bool = envs_ascend.VLLM_ENABLE_MC2
USING_LCCL_COM: bool = envs_ascend.USING_LCCL_COM
@@ -956,6 +958,10 @@ class AscendUnquantizedFusedMoEMethod(UnquantizedFusedMoEMethod):
class AscendFusedMoE(FusedMoE):
# The moe_counter parameter is required during the initialization of EPLB
# to identify the current layer index within the MOE model.
moe_counter = -1
def __init__(
self,
num_experts: int, # Global number of experts
@@ -983,6 +989,9 @@ class AscendFusedMoE(FusedMoE):
# fixme and make __init__() of AscendFusedMoE more clear
super(FusedMoE, self).__init__()
AscendFusedMoE.moe_counter += 1
self.moe_instance_id = AscendFusedMoE.moe_counter
if params_dtype is None:
params_dtype = torch.get_default_dtype()
@@ -1016,16 +1025,33 @@ class AscendFusedMoE(FusedMoE):
self.e_score_correction_bias = e_score_correction_bias
self.expert_map = None
self.activation = activation
self.log2phy = None
self.global_redundant_expert_num = 0
# Create a tensor of size num_experts filled with -1
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size,
get_ep_group().rank_in_group, self.global_num_experts)
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()
else:
# Create a tensor of size num_experts filled with -1
self.local_num_experts, self.expert_map = determine_expert_map(
self.ep_size,
get_ep_group().rank_in_group, self.global_num_experts)
self.moe_parallel_config.tp_rank = get_etp_group().rank_in_group
self.moe_parallel_config.ep_rank = get_ep_group().rank_in_group
ascend_config = get_ascend_config()
self.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
# NOTE: multistream only effective when `VLLM_ENABLE_MC2` is on
self.enable_multistream_shared_expert = \
@@ -1122,6 +1148,8 @@ class AscendFusedMoE(FusedMoE):
e_score_correction_bias=self.e_score_correction_bias,
is_prefill=is_prefill,
enable_force_load_balance=enable_force_load_balance,
log2phy=self.log2phy,
global_redundant_expert_num=self.global_redundant_expert_num,
**kwargs)
if self.enable_multistream_shared_expert and not is_prefill: