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

29
vllm_ascend/eplb/core/eplb_utils.py
View File

@@ -25,32 +25,6 @@ from vllm.logger import logger
import vllm_ascend.envs as envs_ascend
def determine_default_expert_map(global_expert_num, world_size, rank_id,
global_redundant_expert_num):
if world_size == 1:
local_ids = torch.arange(global_expert_num, dtype=torch.int32)
return (global_expert_num, local_ids)
local_num_experts = global_expert_num // world_size
expert_map = torch.full((global_expert_num, ), -1, dtype=torch.int32)
if rank_id < world_size - 1:
start = rank_id * local_num_experts
end = (rank_id + 1) * local_num_experts
local_count = local_num_experts
else:
start = rank_id * local_num_experts
end = global_expert_num
local_count = global_expert_num - rank_id * local_num_experts
if isinstance(local_count, int):
local_ids = torch.arange(local_count, dtype=torch.int32)
expert_map[start:end] = local_ids
return (local_count, expert_map)
def generate_log2phy_map(expert_map):
num_local_experts = expert_map.max() + 1
log2phy_map = expert_map.clone()
@@ -90,8 +64,7 @@ def generate_log2phy_map(expert_map):
return log2phy_map
def determine_default_log2phy_map(global_expert_num, world_size, rank_id,
global_redundant_expert_num):
def determine_default_log2phy_map(global_expert_num, world_size, rank_id):
if world_size == 1:
local_ids = torch.arange(global_expert_num, dtype=torch.int32)
expert_map_all = local_ids.unsqueeze(0).expand(world_size, -1)