ジョブスケジューラのエラー

大学のスパコンはジョブ管理システム PBSによってジョブスケジュールを行なっているのですが，
最近変わったのか，Pythonで並列処理をしようとするとなぜかエラーが出るようになりました。

エラー

=>> PBS: job killed: ncpus 20.8 exceeded limit 16

1	=>> PBS: job killed: ncpus 20.8 exceeded limit 16

コード


import os
from mpi4py import MPI
comm = MPI.COMM_WORLD
size = comm.Get_size()
rank = comm.Get_rank()
name = MPI.Get_processor_name()

import csv
source=[]
target=[]
with open("test.csv","r",encoding="utf-8") as f:
    cdata=csv.reader(f,delimiter=",")
    for i,items in enumerate(cdata):
        if i==0:
            continue
        source.append(items[0])
        target.append(items[1])

data_lst=[]
for src,tgt in zip(source,target):
    data_lst.append((src,tgt))

# http://muscle199x.blog.fc2.com/blog-entry-70.html?sp
def split_list(lst, n):
    list_size = len(lst)
    a = list_size // n
    b = list_size % n
    return [lst[i * a + (i if i < b else b):(i + 1) * a + (i + 1 if i < b else b)] for i in range(n)]

from transformers import MT5ForConditionalGeneration, MT5Model, MT5EncoderModel, T5Tokenizer

tokenizer = T5Tokenizer.from_pretrained("google/mt5-base")
model = MT5ForConditionalGeneration.from_pretrained("google/mt5-base")
if rank == 0:
    data = split_list(data_lst, size)
else:
    data = None

def calc(my_data):
    res=[]

    for src,tgt in my_data:
        inputs = tokenizer.encode("summarize: " + src, return_tensors='pt', max_length=512, truncation=True)
        summary_ids = model.generate(inputs, max_length=50, min_length=10, length_penalty=5., num_beams=2)
        summary = tokenizer.decode(summary_ids[0])
        res.append((src,tgt,summary))
    return res

my_data = comm.scatter(data, root=0)

for r in range(comm.size):
    if rank == r:
        print("scatter.[%d] %d" % (rank, len(my_data)))
    comm.Barrier()

my_res = calc(my_data)
def flatten(l):
    try:
        return ([item for sublist in l for item in sublist])
    except:
        import traceback
        traceback.print_exc()
        print(l)

# gather
res = MPI.COMM_WORLD.gather(my_res, root=0)

if rank == 0:
    res=flatten(res)
    for src,tgt,summary in res:
        print(f"source:{src}")
        print(f"target:{tgt}")
        print(f"summary:{summary}")
        print("-----")

import os

from mpi4py import MPI

comm = MPI.COMM_WORLD

size = comm.Get_size()

rank = comm.Get_rank()

name = MPI.Get_processor_name()

import csv

source=[]

target=[]

with open("test.csv","r",encoding="utf-8") as f:

cdata=csv.reader(f,delimiter=",")

for i,items in enumerate(cdata):

if i==0:

continue

source.append(items[0])

target.append(items[1])

data_lst=[]

for src,tgt in zip(source,target):

data_lst.append((src,tgt))

# http://muscle199x.blog.fc2.com/blog-entry-70.html?sp

def split_list(lst, n):

list_size = len(lst)

a = list_size // n

b = list_size % n

return [lst[i * a + (i if i < b else b):(i + 1) * a + (i + 1 if i < b else b)] for i in range(n)]

from transformers import MT5ForConditionalGeneration, MT5Model, MT5EncoderModel, T5Tokenizer

tokenizer = T5Tokenizer.from_pretrained("google/mt5-base")

model = MT5ForConditionalGeneration.from_pretrained("google/mt5-base")

if rank == 0:

data = split_list(data_lst, size)

else:

data = None

def calc(my_data):

res=[]

for src,tgt in my_data:

inputs = tokenizer.encode("summarize: " + src, return_tensors='pt', max_length=512, truncation=True)

summary_ids = model.generate(inputs, max_length=50, min_length=10, length_penalty=5., num_beams=2)

summary = tokenizer.decode(summary_ids[0])

res.append((src,tgt,summary))

return res

my_data = comm.scatter(data, root=0)

for r in range(comm.size):

if rank == r:

print("scatter.[%d] %d" % (rank, len(my_data)))

comm.Barrier()

my_res = calc(my_data)

def flatten(l):

try:

return ([item for sublist in l for item in sublist])

except:

import traceback

traceback.print_exc()

print(l)

# gather

res = MPI.COMM_WORLD.gather(my_res, root=0)

if rank == 0:

res=flatten(res)

for src,tgt,summary in res:

print(f"source:{src}")

print(f"target:{tgt}")

print(f"summary:{summary}")

print("-----")

これで実行スクリプトはこんな感じ

#!/usr/bin/bash
#PBS -N sample
#PBS -j oe 
#PBS -l select=1:ncpus=64:mpiprocs=64
#PBS -q SINGLE
NPROCS=`cat $PBS_NODEFILE|wc -l`
cd ${PBS_O_WORKDIR}
mpirun -np ${NPROCS} python mt5.py

#!/usr/bin/bash

#PBS -N sample

#PBS -j oe

#PBS -l select=1:ncpus=64:mpiprocs=64

#PBS -q SINGLE

NPROCS=`cat $PBS_NODEFILE|wc -l`

cd ${PBS_O_WORKDIR}

mpirun -np ${NPROCS} python mt5.py

解決方法

ここにありました。

簡単に解決するためには，Pythonコードの一番上に


import os
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"
os.environ["OMP_NUM_THREADS"] = "1"

import os

os.environ["MKL_NUM_THREADS"] = "1"

os.environ["NUMEXPR_NUM_THREADS"] = "1"

os.environ["OMP_NUM_THREADS"] = "1"

これを入れるだけ。Numpyを呼ぶ前に呼ぶ必要があるので，シェルスクリプトで呼んでもいいですし，Pythonの一番上で環境変数に設定してやってもいいです。