隨著AI、大模型的快速發展，傳統的集中式計算已無法應對激增的數據處理需求，而分布式計算是指將一個計算任務分解成多個子任務，由多個計算節點并行地進行計算，并將結果匯總得到最終結果的計算方式，能夠更高效、更穩定、更靈活地處理大規模數據和復雜計算任務，在各行各業中得到了廣泛的應用。

那如何從零到一搭建分布式計算的環境呢？本文將從硬件選型，到服務器側的基礎配置、GPU驅動安裝和集合通訊庫配置，以及無損以太網的啟用，直至大模型導入和訓練測試，帶您跑通搭建分布式計算環境的全流程。

硬件準備

GPU服務器選型

GPU擁有大量的計算核心，可以同時處理多個數據任務，是構成智算中心的關鍵硬件。

從智算中心方案的整體設計層面來看：GPU服務器集群和存儲服務器集群分別通過計算網絡（Scale-out網絡）和存儲網絡連接。另外兩張管理網中，業務管理網用于GPU服務器互聯，進行AIOS管理面通信，帶外管理則連接整個智算中心的所有設備，用于運維接入管理。

明確了智算中心的整體設計后，我們將對比通用計算服務器與GPU服務器的內部硬件連接拓撲圖，來具體了解GPU服務器的選型邏輯：

圖2是一臺通用計算服務器內部的硬件連接拓撲，這臺服務器的核心是兩塊AMD的EPYC CPU，根據IO Chiplet擴展出了若干接口，輔助CPU充分釋放通用計算能力。

圖3是一臺GPU服務器內部的硬件連接拓撲，這臺服務器配備了8塊A100 GPU，8張用于計算通信的RDMA網卡，以及2張用于存儲通信的RDMA網卡，所有的IO組件設計，都是為了讓這8塊GPU充分釋放算力。

通過上面兩張硬件連接拓撲圖可以看到，通用服務器和GPU服務器從基本的硬件構造上就有著非常大的差異，一個是圍繞通用CPU來構建，另一個是圍繞著GPU來構建的。因此，在硬件選型階段，就需要注意差別，通常來講通用服務器是沒有辦法復用改造成一臺高性能的GPU服務器，PCIe接口數量、服務器空間、散熱設計、電源等方面都不能滿足要求。

當通過計算任務確定算力需求，進而確定了所需要的GPU型號和數量之后，我們也就可以再繼續規劃整個GPU集群的組網了。

由于資源限制，本次實驗驗證中，使用三臺通用服務器稍加改造進行后續的并行訓練和推理測試。

計算節點的硬件配置如下：

CPU：Intel(R) Xeon(R) CPU E5-2678 v3 @ 2.50GHz * 2

GPU：NVIDIA GeForce RTX 4060 Ti 16G * 1

內存：128G

存儲：10T HDD * 2

網卡：MGMT、CX5

其他部分：

散熱：GPU為全高尺寸，但服務器只有2U，所以只能拆掉上蓋板;

電源：通用服務器通常沒有預留足夠的供電接口，因此需要使用外置電源對GPU進行額外供電；

電源選擇的是Great Wall 額定650W X6，功率上可以同時滿足3塊GPU（RTX4060Ti需要外接150W的供電）的供電要求，并且支持3個8pin接口，用來分別連接三塊GPU。

高性能計算網選型

智算中心的管理網相較于傳統的通用計算數據中心來說，沒有太大差異。比較特殊的就是Scale-out計算網絡和存儲網絡，這兩張網絡承載的業務流量決定了交換機設備的選型需求：支持RDMA、低時延、高吞吐。

如下圖所示，在組網連接方面也有所不同，這里會通過將GPU分組（圖中#L0～7一組，#L8～15一組），組成只有一跳的高帶寬互聯域（HB域），并通過針對智算場景優化的Rail交換機連接，實現了高效的數據傳輸和計算協同。

這次實驗驗證中，計算網的交換機選用星融元Asterfusion?? CX-N系列超低時延交換機，具體型號為CX308P-48Y-N。

表1：具體型號規格示意

提升大模型訓練效率

CX-N數據中心交換機的單機轉發時延（400ns）低至業界平均水平的1/4~1/5，將網絡時延在AI/ML應用端到端時延中的占比降至最低，同時多維度的高可靠設計確保網絡在任何時候都不中斷，幫助大模型的訓練大幅度降低訓練時間、提升整體效率。

全系列標配RoCEv2能力

區別于傳統廠家多等級License權限管理方式，CX-N數據中心交換機所有應用場景License權限一致，全系列標配RoCEv2能力，提供PFC、ECN、Easy RoCE等一系列面向生產環境的增強網絡特性，用戶無須為此類高級特性額外付出網絡建設成本，幫助用戶獲得更高的ROI。

開放、中立的AI/ML網絡

星融元AI/ML網絡解決方案的開放性確保用戶能夠重用已有的系統（K8s、Prometheus等）對網絡進行管理，無需重復投入；星融元以“中立的網絡供應商參與AI生態”的理念為用戶提供專業的網絡方案，幫助用戶規避“全棧方案鎖定”的風險。

最終，實驗環節的組網拓撲和基礎配置如下所示。

軟件準備

以上，我們已經完成了硬件選型，接下來我們將進行軟件層面的配置：部署 RoCEv2 交換機、配置GPU 服務器、安裝 GPU 驅動和集合通訊庫。

RoCEv2交換機

本次并行訓練的環境中設備數量較少，組網相對簡單：

1. 將CX5網卡的25GE業務接口連接到CX308P；

2. 在交換機上一鍵啟用全局RoCE的無損配置；

3. 將三個25G業務口劃分到一個VLAN下組成一個二層網絡；

如前文提到，CX-N數據中心交換機全系列標配RoCEv2能力，配合星融元AsterNOS網絡操作系統，只需要兩行命令行便可配置所有必要的QoS規則和參數，具體命令行如下：

noone@MacBook-Air ~ % ssh admin@10.230.1.17
Linux AsterNOS 5.10.0-8-2-amd64 #1 SMP Debian 5.10.46-4 (2021-08-03) x86_64
    _          _                _   _   ___   ____  
   /     ___ | |_   ___  _ __ |  | | / _  / ___| 
  / _   / __|| __| / _ | '__||  | || | | |___  
 / ___  __ | |_ |  __/| |   | |  || |_| | ___) |
/_/   _|___/ __| ___||_|   |_| _| ___/ |____/ 

------- Asterfusion Network Operating System -------

Help:    http://www.asterfusion.com/

Last login: Sun Sep 29 17:10:46 2024 from 172.16.20.241

AsterNOS# configure terminal 
AsterNOS(config)# qos roce lossless   
AsterNOS(config)# qos service-policy roce_lossless 
AsterNOS(config)# end
AsterNOS# show qos roce
                    operational    description
------------------  -------------  ---------------------------------------------------
status              bind           qos roce binding status
mode                lossless       Roce Mode
cable-length        40m            Cable Length(in meters) for Roce Lossless Config
congestion-control
- congestion-mode   ECN            congestion-control
- enabled-tc        3,4            Congestion config enabled Traffic Class
- max-threshold     750000         Congestion config max-threshold
- min-threshold     15360          Congestion config max-threshold
pfc
- pfc-priority      3,4            switch-prio on which PFC is enabled
- rx-enabled        enable         PFC Rx Enabled status
- tx-enabled        enable         PFC Tx Enabled status
trust
- trust-mode        dscp           Trust Setting on the port for packet classification

 RoCE DSCP->SP mapping configurations
==========================================
dscp                       switch-prio
-----------------------  -------------
0,1,2,3,4,5,6,7                      0
10,11,12,13,14,15,8,9                1
16,17,18,19,20,21,22,23              2
24,25,26,27,28,29,30,31              3
32,33,34,35,36,37,38,39              4
40,41,42,43,44,45,46,47              5
48,49,50,51,52,53,54,55              6
56,57,58,59,60,61,62,63              7

 RoCE SP->TC mapping and ETS configurations
================================================
  switch-prio  mode    weight
-------------  ------  --------
            6  SP
            7  SP

 RoCE pool config
======================
name                     switch-prio
-----------------------  -------------
egress_lossy_profile     0 1 2 5 6
ingress_lossy_profile    0 1 2 5 6
egress_lossless_profile  3 4
roce_lossless_profile    3 4

GPU服務器基礎配置

以下所有操作，在三臺服務器上都需要執行，本文檔中的配置步驟以server3為例。

關閉防火墻和SELinux

[root@server3 ~]# systemctl stop firewalld
[root@server3 ~]# systemctl disable firewalld
[root@server3 ~]# setenforce 0
[root@server3 ~]# sed -i 's/SELINUX=enforcing/SELINUX=disabled/g' /etc/sysconfig/selinux

配置服務器間免密登陸

[root@server3 ~]# ssh-keygen
[root@server3 ~]# ssh-copy-id root@server1
[root@server3 ~]# ssh-copy-id root@server2

配置服務器軟件源

[root@server3 ~]# ll /etc/yum.repos.d/
總用量 80
-rw-r--r-- 1 root root 2278 9月  19 08:00 CentOS-Base.repo
-rw-r--r-- 1 root root  232 9月  19 08:00 cuda-rhel7.repo
-rw-r--r-- 1 root root  210 9月  19 08:00 cudnn-local-rhel7-8.9.7.29.repo
drwxr-xr-x 2 root root 4096 9月  19 07:58 disable.d
-rw-r--r-- 1 root root  664 9月  19 08:00 epel.repo
-rw-r--r-- 1 root root  381 9月  19 08:00 hashicorp.repo
-rw-r--r-- 1 root root  218 9月  19 08:00 kubernetes.repo
-rw-r--r-- 1 root root  152 9月  19 08:00 MariaDB.repo
-rw-r--r-- 1 root root  855 9月  19 08:00 remi-modular.repo
-rw-r--r-- 1 root root  456 9月  19 08:00 remi-php54.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php70.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php71.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php72.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php73.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php74.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php80.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php81.repo
-rw-r--r-- 1 root root 1314 9月  19 08:00 remi-php82.repo
-rw-r--r-- 1 root root 2605 9月  19 08:00 remi.repo
-rw-r--r-- 1 root root  750 9月  19 08:00 remi-safe.repo
[root@server3 ~]# more /etc/yum.repos.d/*.repo
::::::::::::::
/etc/yum.repos.d/CentOS-Base.repo
::::::::::::::
# CentOS-Base.repo
#
# The mirror system uses the connecting IP address of the client and the
# update status of each mirror to pick mirrors that are updated to and
# geographically close to the client.  You should use this for CentOS updates
# unless you are manually picking other mirrors.
#
# If the mirrorlist= does not work for you, as a fall back you can try the 
# remarked out baseurl= line instead.
#
#
 
[base]
name=CentOS-7 - Base - mirrors.aliyun.com
failovermethod=priority
baseurl=http://mirrors.aliyun.com/centos/7/os/x86_64/
        http://mirrors.aliyuncs.com/centos/7/os/x86_64/
        http://mirrors.cloud.aliyuncs.com/centos/7/os/x86_64/
gpgcheck=1
gpgkey=http://mirrors.aliyun.com/centos/RPM-GPG-KEY-CentOS-7
 
#released updates 
[updates]
name=CentOS-7 - Updates - mirrors.aliyun.com
failovermethod=priority
baseurl=http://mirrors.aliyun.com/centos/7/updates/x86_64/
        http://mirrors.aliyuncs.com/centos/7/updates/x86_64/
        http://mirrors.cloud.aliyuncs.com/centos/7/updates/x86_64/
gpgcheck=1
gpgkey=http://mirrors.aliyun.com/centos/RPM-GPG-KEY-CentOS-7
 
#additional packages that may be useful
[extras]
name=CentOS-7 - Extras - mirrors.aliyun.com
failovermethod=priority
baseurl=http://mirrors.aliyun.com/centos/7/extras/x86_64/
        http://mirrors.aliyuncs.com/centos/7/extras/x86_64/
        http://mirrors.cloud.aliyuncs.com/centos/7/extras/x86_64/
gpgcheck=1
gpgkey=http://mirrors.aliyun.com/centos/RPM-GPG-KEY-CentOS-7
 
#additional packages that extend functionality of existing packages
[centosplus]
name=CentOS-7 - Plus - mirrors.aliyun.com
failovermethod=priority
baseurl=http://mirrors.aliyun.com/centos/7/centosplus/x86_64/
        http://mirrors.aliyuncs.com/centos/7/centosplus/x86_64/
        http://mirrors.cloud.aliyuncs.com/centos/7/centosplus/x86_64/
gpgcheck=1
enabled=0
gpgkey=http://mirrors.aliyun.com/centos/RPM-GPG-KEY-CentOS-7
 
#contrib - packages by Centos Users
[contrib]
name=CentOS-7 - Contrib - mirrors.aliyun.com
failovermethod=priority
baseurl=http://mirrors.aliyun.com/centos/7/contrib/x86_64/
        http://mirrors.aliyuncs.com/centos/7/contrib/x86_64/
        http://mirrors.cloud.aliyuncs.com/centos/7/contrib/x86_64/
gpgcheck=1
enabled=0
gpgkey=http://mirrors.aliyun.com/centos/RPM-GPG-KEY-CentOS-7
::::::::::::::
/etc/yum.repos.d/cuda-rhel7.repo
::::::::::::::
[cuda-rhel7-x86_64]
name=cuda-rhel7-x86_64
baseurl=https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64
enabled=1
gpgcheck=1
gpgkey=https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/D42D0685.pub
::::::::::::::
/etc/yum.repos.d/cudnn-local-rhel7-8.9.7.29.repo
::::::::::::::
[cudnn-local-rhel7-8.9.7.29]
name=cudnn-local-rhel7-8.9.7.29
baseurl=file:///var/cudnn-local-repo-rhel7-8.9.7.29
enabled=1
gpgcheck=1
gpgkey=file:///var/cudnn-local-repo-rhel7-8.9.7.29/90F10142.pub
obsoletes=0
::::::::::::::
/etc/yum.repos.d/epel.repo
::::::::::::::
[epel]
name=Extra Packages for Enterprise Linux 7 - $basearch
baseurl=http://mirrors.aliyun.com/epel/7/$basearch
failovermethod=priority
enabled=1
gpgcheck=0
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-EPEL-7
 
[epel-debuginfo]
name=Extra Packages for Enterprise Linux 7 - $basearch - Debug
baseurl=http://mirrors.aliyun.com/epel/7/$basearch/debug
failovermethod=priority
enabled=0
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-EPEL-7
gpgcheck=0
 
[epel-source]
name=Extra Packages for Enterprise Linux 7 - $basearch - Source
baseurl=http://mirrors.aliyun.com/epel/7/SRPMS
failovermethod=priority
enabled=0
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-EPEL-7
gpgcheck=0
::::::::::::::
/etc/yum.repos.d/hashicorp.repo
::::::::::::::
[hashicorp]
name=Hashicorp Stable - $basearch
baseurl=https://rpm.releases.hashicorp.com/RHEL/$releasever/$basearch/stable
enabled=0
gpgcheck=1
gpgkey=https://rpm.releases.hashicorp.com/gpg

[hashicorp-test]
name=Hashicorp Test - $basearch
baseurl=https://rpm.releases.hashicorp.com/RHEL/$releasever/$basearch/test
enabled=0
gpgcheck=1
gpgkey=https://rpm.releases.hashicorp.com/gpg
::::::::::::::
/etc/yum.repos.d/kubernetes.repo
::::::::::::::
[kubernetes]
name=Kubernetes
baseurl=https://mirrors.aliyun.com/kubernetes-new/core/stable/v1.28/rpm/
enabled=1
gpgcheck=1
gpgkey=https://mirrors.aliyun.com/kubernetes-new/core/stable/v1.28/rpm/repodata/repomd.xml.key
::::::::::::::
/etc/yum.repos.d/MariaDB.repo
::::::::::::::
[mariadb]
name = MariaDB
baseurl = https://mirror.mariadb.org/yum/11.2/centos74-amd64
gpgkey = https://yum.mariadb.org/RPM-GPG-KEY-MariaDB
gpgcheck = 0
::::::::::::::
/etc/yum.repos.d/remi-modular.repo
::::::::::::::
# Repository: https://rpms.remirepo.net/
# Blog:       https://blog.remirepo.net/
# Forum:      https://forum.remirepo.net/

[remi-modular]
name=Remi's Modular repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/modular/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/modular/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/modular/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-modular-test]
name=Remi's Modular testing repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/modular-test/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/modular-test/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/modular-test/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

::::::::::::::
/etc/yum.repos.d/remi-php54.repo
::::::::::::::
# This repository only provides PHP 5.4 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php54]
name=Remi's PHP 5.4 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php54/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php54/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php54/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

::::::::::::::
/etc/yum.repos.d/remi-php70.repo
::::::::::::::
# This repository only provides PHP 7.0 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php70]
name=Remi's PHP 7.0 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php70/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php70/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php70/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php70-debuginfo]
name=Remi's PHP 7.0 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php70/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php70-test]
name=Remi's PHP 7.0 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test70/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test70/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test70/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php70-test-debuginfo]
name=Remi's PHP 7.0 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test70/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php71.repo
::::::::::::::
# This repository only provides PHP 7.1 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php71]
name=Remi's PHP 7.1 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php71/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php71/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php71/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php71-debuginfo]
name=Remi's PHP 7.1 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php71/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php71-test]
name=Remi's PHP 7.1 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test71/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test71/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test71/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php71-test-debuginfo]
name=Remi's PHP 7.1 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test71/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php72.repo
::::::::::::::
# This repository only provides PHP 7.2 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php72]
name=Remi's PHP 7.2 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php72/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php72/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php72/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php72-debuginfo]
name=Remi's PHP 7.2 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php72/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php72-test]
name=Remi's PHP 7.2 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test72/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test72/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test72/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php72-test-debuginfo]
name=Remi's PHP 7.2 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test72/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php73.repo
::::::::::::::
# This repository only provides PHP 7.3 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php73]
name=Remi's PHP 7.3 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php73/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php73/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php73/mirror
enabled=1
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php73-debuginfo]
name=Remi's PHP 7.3 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php73/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php73-test]
name=Remi's PHP 7.3 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test73/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test73/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test73/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php73-test-debuginfo]
name=Remi's PHP 7.3 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test73/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php74.repo
::::::::::::::
# This repository only provides PHP 7.4 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php74]
name=Remi's PHP 7.4 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php74/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php74/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php74/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php74-debuginfo]
name=Remi's PHP 7.4 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php74/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php74-test]
name=Remi's PHP 7.4 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test74/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test74/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test74/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php74-test-debuginfo]
name=Remi's PHP 7.4 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test74/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php80.repo
::::::::::::::
# This repository only provides PHP 8.0 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php80]
name=Remi's PHP 8.0 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php80/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php80/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php80/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php80-debuginfo]
name=Remi's PHP 8.0 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php80/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php80-test]
name=Remi's PHP 8.0 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test80/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test80/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test80/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php80-test-debuginfo]
name=Remi's PHP 8.0 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test80/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php81.repo
::::::::::::::
# This repository only provides PHP 8.1 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php81]
name=Remi's PHP 8.1 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php81/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php81/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php81/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php81-debuginfo]
name=Remi's PHP 8.1 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php81/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php81-test]
name=Remi's PHP 8.1 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test81/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test81/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test81/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php81-test-debuginfo]
name=Remi's PHP 8.1 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test81/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi-php82.repo
::::::::::::::
# This repository only provides PHP 8.2 and its extensions
# NOTICE: common dependencies are in "remi-safe"

[remi-php82]
name=Remi's PHP 8.2 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php82/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php82/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php82/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php82-debuginfo]
name=Remi's PHP 8.2 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php82/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php82-test]
name=Remi's PHP 8.2 test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test82/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test82/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test82/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php82-test-debuginfo]
name=Remi's PHP 8.2 test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test82/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
::::::::::::::
/etc/yum.repos.d/remi.repo
::::::::::::::
# Repository: http://rpms.remirepo.net/
# Blog:       http://blog.remirepo.net/
# Forum:      http://forum.remirepo.net/

[remi]
name=Remi's RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/remi/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/remi/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/remi/mirror
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php55]
name=Remi's PHP 5.5 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php55/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php55/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php55/mirror
# NOTICE: common dependencies are in "remi-safe"
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php56]
name=Remi's PHP 5.6 RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/php56/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/php56/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/php56/mirror
# NOTICE: common dependencies are in "remi-safe"
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-test]
name=Remi's test RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/test/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/test/mirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/test/mirror
# WARNING: If you enable this repository, you must also enable "remi"
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-debuginfo]
name=Remi's RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-remi/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php55-debuginfo]
name=Remi's PHP 5.5 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php55/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-php56-debuginfo]
name=Remi's PHP 5.6 RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-php56/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-test-debuginfo]
name=Remi's test RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-test/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

::::::::::::::
/etc/yum.repos.d/remi-safe.repo
::::::::::::::
# This repository is safe to use with RHEL/CentOS base repository
# it only provides additional packages for the PHP stack
# all dependencies are in base repository or in EPEL

[remi-safe]
name=Safe Remi's RPM repository for Enterprise Linux 7 - $basearch
#baseurl=http://rpms.remirepo.net/enterprise/7/safe/$basearch/
#mirrorlist=https://rpms.remirepo.net/enterprise/7/safe/httpsmirror
mirrorlist=http://cdn.remirepo.net/enterprise/7/safe/mirror
enabled=1
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi

[remi-safe-debuginfo]
name=Remi's RPM repository for Enterprise Linux 7 - $basearch - debuginfo
baseurl=http://rpms.remirepo.net/enterprise/7/debug-remi/$basearch/
enabled=0
gpgcheck=1
gpgkey=file:///etc/pki/rpm-gpg/RPM-GPG-KEY-remi
[root@server3 ~]# 

安裝Python3

準備工作目錄
[root@server3 lichao]# mkdir AIGC
[root@server3 lichao]# cd AIGC/

安裝Python3

安裝編譯環境和依賴包
[root@server3 AIGC]# yum install wget gcc openssl-devel bzip2-devel libffi-devel
[root@server3 AIGC]# yum install openssl11 openssl11-devel openssl-devel
解壓源碼包
[root@server3 AIGC]# tar xvf Python-3.11.9.tar.xz 
[root@server3 AIGC]# cd Python-3.11.9
[root@server3 Python-3.11.9]# 
設置環境變量
[root@server3 Python-3.11.9]# export CFLAGS=$(pkg-config --cflags openssl11)
[root@server3 Python-3.11.9]# export LDFLAGS=$(pkg-config --libs openssl11)
進行編譯安裝
[root@server3 Python-3.11.9]# mkdir -p /home/lichao/opt/python3.11.9
[root@server3 Python-3.11.9]# ./configure --prefix=/home/lichao/opt/python3.11.9
[root@server3 Python-3.11.9]# make && make install
創建軟鏈接，用于全局訪問
[root@server3 Python-3.11.9]# cd /home/lichao/opt/python3.11.9/
[root@server3 python3.11.9]# ln -s /home/lichao/opt/python3.11.9/bin/python3 /usr/bin/python3
[root@server3 python3.11.9]# ln -s /home/lichao/opt/python3.11.9/bin/pip3 /usr/bin/pip3
[root@server3 python3.11.9]# ll /usr/bin/python3 
lrwxrwxrwx 1 root root 41 5月  16 08:32 /usr/bin/python3 -> /home/lichao/opt/python3.11.9/bin/python3
[root@server3 python3.11.9]# ll /usr/bin/pip3
lrwxrwxrwx 1 root root 38 5月  16 08:32 /usr/bin/pip3 -> /home/lichao/opt/python3.11.9/bin/pip3
驗證測試
[root@server3 python3.11.9]# python3
Python 3.11.9 (main, May 16 2024, 08:23:00) [GCC 4.8.5 20150623 (Red Hat 4.8.5-44)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> exit()
[root@server3 python3.11.9]# 

安裝MLNX網卡驅動

下文以CentOS7為例，詳細介紹了Mellanox網卡MLNX_OFED的驅動安裝和固件升級方法。

本次下載的驅動版本為：MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64.tgz。

[root@server3 ~]# tar –zxvf MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64.tgz
[root@server3 ~]# cd MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64
查看當前系統的內核版本
[root@server3 MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64]# uname -r
3.10.0-957.el7.x86_64
查看當前驅動所支持的內核版本
[root@server3 MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64]# cat .supported_kernels 
3.10.0-957.el7.x86_64 
注：由以上可知下載的默認驅動支持當前的內核版本
如果當前內核與支持內核不匹配，手動編譯適合內核的驅動，在編譯之前首先安裝gcc編譯環境和kernel開發包
[root@server3 MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64]#yum  install gcc gcc-c++
libstdc++-devel kernel-default-devel 
添加針對當前內核版本的驅動
[root@server3 MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64]#./mlnx_add_kernel_support.sh -m /root/MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64  -v
注：完成后生成的驅動文件在/tmp目錄下
[root@server3 MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64]# ls -l /tmp/MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64-ext.tgz
-rw-r--r-- 1 root root 282193833 Dec 23 09:49 /tmp/MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64-ext.tgz
安裝驅動
[root@server3 tmp]# tar xzvf MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64-ext.tgz
[root@server3 tmp]# cd MLNX_OFED_LINUX-4.7-3.2.9.0-rhel7.6-x86_64-ext
[root@server3 tmp]# ./mlnxofedinstall
最后啟動openibd服務
[root@server3 ~]#/etc/init.d/openibd start
[root@server3 ~]#chkconfig openibd on

安裝GPU驅動和集合通訊庫安裝配置

安裝配置

• 安裝GPU驅動和CUDA、CUDNN

安裝開始前，請根據自己的GPU型號、操作系統版本去英偉達官網下載相對應的軟件包。

[root@server3 AIGC]# ll
總用量 1733448
-rw-r--r--  1 root root 1430373861 5月  16 08:55 cudnn-local-repo-rhel7-8.9.7.29-1.0-1.x86_64.rpm
drwxr-xr-x  7 root root        141 5月  17 13:45 nccl-tests
-rwxr-xr-x  1 root root  306736632 5月  16 08:43 NVIDIA-Linux-x86_64-550.67.run
drwxrwxr-x 10 1000 1000       4096 5月  17 13:21 openmpi-4.1.6
-rw-r--r--  1 root root   17751702 9月  30 2023 openmpi-4.1.6.tar.gz
drwxr-xr-x 17 root root       4096 5月  16 08:23 Python-3.11.9
-rw-r--r--  1 root root   20175816 4月   2 13:11 Python-3.11.9.tar.xz
[root@server3 AIGC]# ./NVIDIA-Linux-x86_64-550.67.run
Verifying archive integrity... OK
Uncompressing NVIDIA Accelerated Graphics Driver for Linux-x86_64 550.67...................

[root@server3 AIGC]# yum-config-manager --add-repo https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
已加載插件：fastestmirror, nvidia
adding repo from: https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo
grabbing file https://developer.download.nvidia.com/compute/cuda/repos/rhel7/x86_64/cuda-rhel7.repo to /etc/yum.repos.d/cuda-rhel7.repo
repo saved to /etc/yum.repos.d/cuda-rhel7.repo
[root@server3 AIGC]# yum install libnccl-2.21.5-1+cuda12.4 libnccl-devel-2.21.5-1+cuda12.4 libnccl-static-2.21.5-1+cuda12.4
[root@server3 AIGC]# yum install cudnn-local-repo-rhel7-8.9.7.29-1.0-1.x86_64.rpm

安裝完成后，可以通過nvidia-smi查看驅動和CUDA版本。如果版本不匹配，則執行此命令行會報錯。

[root@server3 AIGC]# nvidia-smi 
Mon Jun  3 11:59:36 2024       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 550.67                 Driver Version: 550.67         CUDA Version: 12.4     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4060 Ti     Off |   00000000:02:00.0 Off |                  N/A |
|  0%   34C    P0             27W /  165W |       1MiB /  16380MiB |      0%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+
                                                                                         
+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI        PID   Type   Process name                              GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
[root@server3 AIGC]# 

• 編譯安裝OpenMPI

[root@server3 AIGC]# tar xvf openmpi-4.1.6.tar.gz 
[root@server3 openmpi-4.1.6]# 
[root@server3 openmpi-4.1.6]# mkdir -p /home/lichao/lib/openmpi
[root@server3 openmpi-4.1.6]# ./configure --prefix=/home/lichao/lib/openmpi -with-cuda=/usr/local/cuda-12.4 -with-nccl=/usr/lib64

Open MPI configuration:
-----------------------
Version: 4.1.6
Build MPI C bindings: yes
Build MPI C++ bindings (deprecated): no
Build MPI Fortran bindings: mpif.h, use mpi
MPI Build Java bindings (experimental): no
Build Open SHMEM support: yes
Debug build: no
Platform file: (none)

Miscellaneous
-----------------------
CUDA support: yes
HWLOC support: internal
Libevent support: internal
Open UCC: no
PMIx support: Internal
 
Transports
-----------------------
Cisco usNIC: no
Cray uGNI (Gemini/Aries): no
Intel Omnipath (PSM2): no
Intel TrueScale (PSM): no
Mellanox MXM: no
Open UCX: yes
OpenFabrics OFI Libfabric: no
OpenFabrics Verbs: yes
Portals4: no
Shared memory/copy in+copy out: yes
Shared memory/Linux CMA: yes
Shared memory/Linux KNEM: no
Shared memory/XPMEM: no
TCP: yes
 
Resource Managers
-----------------------
Cray Alps: no
Grid Engine: no
LSF: no
Moab: no
Slurm: yes
ssh/rsh: yes
Torque: no
 
OMPIO File Systems
-----------------------
DDN Infinite Memory Engine: no
Generic Unix FS: yes
IBM Spectrum Scale/GPFS: no
Lustre: no
PVFS2/OrangeFS: no
 
[root@server3 openmpi-4.1.6]# 

• 編譯安裝NCCL-Test

[root@server3 lichao]# cd AIGC/
[root@server3 AIGC]# git clone https://github.com/NVIDIA/nccl-tests.git
[root@server3 AIGC]# cd nccl-tests/
[root@server3 nccl-tests]# make clean
[root@server3 nccl-tests]# make MPI=1 MPI_HOME=/home/lichao/opt/openmpi/ CUDA_HOME=/usr/local/cuda-12.4/ NCCL_HOME=/usr/lib64/

集合通信性能測試方法（all_reduce）

[root@server1 lichao]# cat run_nccl-test.sh

/home/lichao/opt/openmpi/bin/mpirun --allow-run-as-root 

-np 3 

-host "server1,server2,server3" 

-mca btl ^openib 

-x NCCL_DEBUG=INFO 

-x NCCL_ALGO=ring 

-x NCCL_IB_DISABLE=0 

-x NCCL_IB_GID_INDEX=3 

-x NCCL_SOCKET_IFNAME=ens11f1 

-x NCCL_IB_HCA=mlx5_1:1 

/home/lichao/AIGC/nccl-tests/build/all_reduce_perf -b 128 -e 8G -f 2 -g 1

[root@server1 lichao]# ./run_nccl-test.sh

# nThread 1 nGpus 1 minBytes 128 maxBytes 8589934592 step: 2(factor) warmup iters: 5 iters: 20 agg iters: 1 validation: 1 graph: 0

#

# Using devices

# Rank 0 Group 0 Pid 18697 on  server1 device 0 [0x02] NVIDIA GeForce RTX 4060 Ti

# Rank 1 Group 0 Pid 20893 on  server2 device 0 [0x02] NVIDIA GeForce RTX 4060 Ti

# Rank 2 Group 0 Pid  2458 on  server3 device 0 [0x02] NVIDIA GeForce RTX 4060 Ti

#

# Reducing maxBytes to 5261099008 due to memory limitation

server1:18697:18697 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server1:18697:18697 [0] NCCL INFO Bootstrap : Using ens11f1:172.16.0.11

server1:18697:18697 [0] NCCL INFO NET/Plugin: No plugin found (libnccl-net.so)

server1:18697:18697 [0] NCCL INFO NET/Plugin: Plugin load returned 2 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-net.so

server1:18697:18697 [0] NCCL INFO NET/Plugin: Using internal network plugin.

server2:20893:20893 [0] NCCL INFO cudaDriverVersion 12040

server2:20893:20893 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server2:20893:20893 [0] NCCL INFO Bootstrap : Using ens11f1:172.16.0.12

server2:20893:20893 [0] NCCL INFO NET/Plugin: No plugin found (libnccl-net.so)

server2:20893:20893 [0] NCCL INFO NET/Plugin: Plugin load returned 2 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-net.so

server2:20893:20893 [0] NCCL INFO NET/Plugin: Using internal network plugin.

server1:18697:18697 [0] NCCL INFO cudaDriverVersion 12040

NCCL version 2.21.5+cuda12.4

server3:2458:2458 [0] NCCL INFO cudaDriverVersion 12040

server3:2458:2458 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server3:2458:2458 [0] NCCL INFO Bootstrap : Using ens11f1:172.16.0.13

server3:2458:2458 [0] NCCL INFO NET/Plugin: No plugin found (libnccl-net.so)

server3:2458:2458 [0] NCCL INFO NET/Plugin: Plugin load returned 2 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-net.so

server3:2458:2458 [0] NCCL INFO NET/Plugin: Using internal network plugin.

server2:20893:20907 [0] NCCL INFO NCCL_IB_DISABLE set by environment to 0.

server2:20893:20907 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server2:20893:20907 [0] NCCL INFO NCCL_IB_HCA set to mlx5_1:1

server2:20893:20907 [0] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [RO]; OOB ens11f1:172.16.0.12

server2:20893:20907 [0] NCCL INFO Using non-device net plugin version 0

server2:20893:20907 [0] NCCL INFO Using network IB

server3:2458:2473 [0] NCCL INFO NCCL_IB_DISABLE set by environment to 0.

server3:2458:2473 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server3:2458:2473 [0] NCCL INFO NCCL_IB_HCA set to mlx5_1:1

server1:18697:18712 [0] NCCL INFO NCCL_IB_DISABLE set by environment to 0.

server1:18697:18712 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to ens11f1

server3:2458:2473 [0] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [RO]; OOB ens11f1:172.16.0.13

server1:18697:18712 [0] NCCL INFO NCCL_IB_HCA set to mlx5_1:1

server3:2458:2473 [0] NCCL INFO Using non-device net plugin version 0

server3:2458:2473 [0] NCCL INFO Using network IB

server1:18697:18712 [0] NCCL INFO NET/IB : Using [0]mlx5_1:1/RoCE [RO]; OOB ens11f1:172.16.0.11

server1:18697:18712 [0] NCCL INFO Using non-device net plugin version 0

server1:18697:18712 [0] NCCL INFO Using network IB

server1:18697:18712 [0] NCCL INFO ncclCommInitRank comm 0x23622c0 rank 0 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init START

server3:2458:2473 [0] NCCL INFO ncclCommInitRank comm 0x346ffc0 rank 2 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init START

server2:20893:20907 [0] NCCL INFO ncclCommInitRank comm 0x2a1af20 rank 1 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init START

server3:2458:2473 [0] NCCL INFO Setting affinity for GPU 0 to 0f,ff000fff

server2:20893:20907 [0] NCCL INFO Setting affinity for GPU 0 to 0f,ff000fff

server1:18697:18712 [0] NCCL INFO Setting affinity for GPU 0 to 0f,ff000fff

server1:18697:18712 [0] NCCL INFO comm 0x23622c0 rank 0 nRanks 3 nNodes 3 localRanks 1 localRank 0 MNNVL 0

server1:18697:18712 [0] NCCL INFO Channel 00/02 :  0  1  2

server1:18697:18712 [0] NCCL INFO Channel 01/02 :  0  1  2

server1:18697:18712 [0] NCCL INFO Trees [0] 2/-1/-1->0->-1 [1] 2/-1/-1->0->1

server1:18697:18712 [0] NCCL INFO P2P Chunksize set to 131072

server3:2458:2473 [0] NCCL INFO comm 0x346ffc0 rank 2 nRanks 3 nNodes 3 localRanks 1 localRank 0 MNNVL 0

server2:20893:20907 [0] NCCL INFO comm 0x2a1af20 rank 1 nRanks 3 nNodes 3 localRanks 1 localRank 0 MNNVL 0

server3:2458:2473 [0] NCCL INFO Trees [0] 1/-1/-1->2->0 [1] -1/-1/-1->2->0

server3:2458:2473 [0] NCCL INFO P2P Chunksize set to 131072

server2:20893:20907 [0] NCCL INFO Trees [0] -1/-1/-1->1->2 [1] 0/-1/-1->1->-1

server2:20893:20907 [0] NCCL INFO P2P Chunksize set to 131072

server3:2458:2473 [0] NCCL INFO Channel 00/0 : 1[0] -> 2[0] [receive] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 01/0 : 1[0] -> 2[0] [receive] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 00/0 : 2[0] -> 0[0] [send] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 01/0 : 2[0] -> 0[0] [send] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[0] [receive] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[0] [receive] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 00/0 : 1[0] -> 2[0] [send] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 01/0 : 1[0] -> 2[0] [send] via NET/IB/0

server1:18697:18712 [0] NCCL INFO Channel 00/0 : 2[0] -> 0[0] [receive] via NET/IB/0

server1:18697:18712 [0] NCCL INFO Channel 01/0 : 2[0] -> 0[0] [receive] via NET/IB/0

server1:18697:18712 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[0] [send] via NET/IB/0

server1:18697:18712 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[0] [send] via NET/IB/0

server3:2458:2475 [0] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3.

server1:18697:18714 [0] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3.

server2:20893:20909 [0] NCCL INFO NCCL_IB_GID_INDEX set by environment to 3.

server1:18697:18712 [0] NCCL INFO Connected all rings

server1:18697:18712 [0] NCCL INFO Channel 01/0 : 1[0] -> 0[0] [receive] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Connected all rings

server2:20893:20907 [0] NCCL INFO Connected all rings

server1:18697:18712 [0] NCCL INFO Channel 00/0 : 0[0] -> 2[0] [send] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 00/0 : 2[0] -> 1[0] [receive] via NET/IB/0

server1:18697:18712 [0] NCCL INFO Channel 01/0 : 0[0] -> 2[0] [send] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 00/0 : 0[0] -> 2[0] [receive] via NET/IB/0

server2:20893:20907 [0] NCCL INFO Channel 01/0 : 1[0] -> 0[0] [send] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 01/0 : 0[0] -> 2[0] [receive] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Channel 00/0 : 2[0] -> 1[0] [send] via NET/IB/0

server3:2458:2473 [0] NCCL INFO Connected all trees

server1:18697:18712 [0] NCCL INFO Connected all trees

server1:18697:18712 [0] NCCL INFO NCCL_ALGO set by environment to ring

server3:2458:2473 [0] NCCL INFO NCCL_ALGO set by environment to ring

server3:2458:2473 [0] NCCL INFO threadThresholds 8/8/64 | 24/8/64 | 512 | 512

server3:2458:2473 [0] NCCL INFO 2 coll channels, 2 collnet channels, 0 nvls channels, 2 p2p channels, 2 p2p channels per peer

server2:20893:20907 [0] NCCL INFO Connected all trees

server2:20893:20907 [0] NCCL INFO NCCL_ALGO set by environment to ring

server2:20893:20907 [0] NCCL INFO threadThresholds 8/8/64 | 24/8/64 | 512 | 512

server2:20893:20907 [0] NCCL INFO 2 coll channels, 2 collnet channels, 0 nvls channels, 2 p2p channels, 2 p2p channels per peer

server1:18697:18712 [0] NCCL INFO threadThresholds 8/8/64 | 24/8/64 | 512 | 512

server1:18697:18712 [0] NCCL INFO 2 coll channels, 2 collnet channels, 0 nvls channels, 2 p2p channels, 2 p2p channels per peer

server2:20893:20907 [0] NCCL INFO TUNER/Plugin: Plugin load returned 11 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-tuner.so

server2:20893:20907 [0] NCCL INFO TUNER/Plugin: Using internal tuner plugin.

server2:20893:20907 [0] NCCL INFO ncclCommInitRank comm 0x2a1af20 rank 1 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init COMPLETE

server3:2458:2473 [0] NCCL INFO TUNER/Plugin: Plugin load returned 11 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-tuner.so

server3:2458:2473 [0] NCCL INFO TUNER/Plugin: Using internal tuner plugin.

server3:2458:2473 [0] NCCL INFO ncclCommInitRank comm 0x346ffc0 rank 2 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init COMPLETE

server1:18697:18712 [0] NCCL INFO TUNER/Plugin: Plugin load returned 11 : libnccl-net.so: cannot open shared object file: No such file or directory : when loading libnccl-tuner.so

server1:18697:18712 [0] NCCL INFO TUNER/Plugin: Using internal tuner plugin.

server1:18697:18712 [0] NCCL INFO ncclCommInitRank comm 0x23622c0 rank 0 nranks 3 cudaDev 0 nvmlDev 0 busId 2000 commId 0x35491327c8228dd0 - Init COMPLETE

#

#                               out-of-place            in-place     

#    size     count   type  redop  root   time  algbw  busbw #wrong   time  algbw  busbw #wrong

#    (B)  (elements)                (us) (GB/s) (GB/s)      (us) (GB/s) (GB/s)   

    128      32   float   sum   -1  28.39  0.00  0.01   0  27.35  0.00  0.01   0

    256      64   float   sum   -1  29.44  0.01  0.01   0  28.54  0.01  0.01   0

    512      128   float   sum   -1  29.99  0.02  0.02   0  29.66  0.02  0.02   0

    1024      256   float   sum   -1  32.89  0.03  0.04   0  30.64  0.03  0.04   0

    2048      512   float   sum   -1  34.81  0.06  0.08   0  31.87  0.06  0.09   0

    4096     1024   float   sum   -1  37.32  0.11  0.15   0  36.09  0.11  0.15   0

    8192     2048   float   sum   -1  45.11  0.18  0.24   0  43.12  0.19  0.25   0

   16384     4096   float   sum   -1  57.92  0.28  0.38   0  56.98  0.29  0.38   0

   32768     8192   float   sum   -1  72.68  0.45  0.60   0  70.79  0.46  0.62   0

   65536     16384   float   sum   -1  95.77  0.68  0.91   0  93.73  0.70  0.93   0

   131072     32768   float   sum   -1  162.7  0.81  1.07   0  161.5  0.81  1.08   0

   262144     65536   float   sum   -1  177.3  1.48  1.97   0  177.4  1.48  1.97   0

   524288    131072   float   sum   -1  301.4  1.74  2.32   0  302.0  1.74  2.31   0

  1048576    262144   float   sum   -1  557.9  1.88  2.51   0  559.2  1.88  2.50   0

  2097152    524288   float   sum   -1  1089.8  1.92  2.57   0  1092.2  1.92  2.56   0

  4194304    1048576   float   sum   -1  2165.7  1.94  2.58   0  2166.6  1.94  2.58   0

  8388608    2097152   float   sum   -1  4315.7  1.94  2.59   0  4316.1  1.94  2.59   0

  16777216    4194304   float   sum   -1  8528.8  1.97  2.62   0  8529.3  1.97  2.62   0

  33554432    8388608   float   sum   -1  16622  2.02  2.69   0  16610  2.02  2.69   0

  67108864   16777216   float   sum   -1  32602  2.06  2.74   0  32542  2.06  2.75   0

 134217728   33554432   float   sum   -1  63946  2.10  2.80   0  63831  2.10  2.80   0

 268435456   67108864   float   sum   -1  126529  2.12  2.83   0  126412  2.12  2.83   0

 536870912   134217728   float   sum   -1  251599  2.13  2.85   0  251327  2.14  2.85   0

 1073741824   268435456   float   sum   -1  500664  2.14  2.86   0  501911  2.14  2.85   0

 2147483648   536870912   float   sum   -1 1001415  2.14  2.86   0 1000178  2.15  2.86   0

 4294967296  1073741824   float   sum   -1 1999361  2.15  2.86   0 1997380  2.15  2.87   0

server1:18697:18697 [0] NCCL INFO comm 0x23622c0 rank 0 nranks 3 cudaDev 0 busId 2000 - Destroy COMPLETE

server2:20893:20893 [0] NCCL INFO comm 0x2a1af20 rank 1 nranks 3 cudaDev 0 busId 2000 - Destroy COMPLETE

server3:2458:2458 [0] NCCL INFO comm 0x346ffc0 rank 2 nranks 3 cudaDev 0 busId 2000 - Destroy COMPLETE

# Out of bounds values : 0 OK

# Avg bus bandwidth  : 1.66163

#



[root@server1 lichao]#

結果詳解

- size (B)：操作處理的數據的大小，以字節為單位；

- count (elements)：操作處理的元素的數量；

- type：元素的數據類型；

- redop：使用的歸約操作；

- root：對于某些操作（如 reduce 和 broadcast），這列指定了根節點的編號，值是 -1 表示這個操作沒有根節點（all-reduce 操作涉及到所有的節點）；

- time (us)：操作的執行時間，以微秒為單位；

- algbw (GB/s)：算法帶寬，以每秒吉字節（GB/s）為單位；

- busbw (GB/s)：總線帶寬，以每秒吉字節（GB/s）為單位；

- wrong：錯誤的數量，如果這個值不是 0，那可能表示有一些錯誤發生。

在這個例子中，你可以看到，當處理的數據量增大時，算法帶寬和總線帶寬都有所提高，這可能表示 NCCL 能夠有效地利用大量的數據。

查看結果時，需要關注如下幾點：

1. 數據量增加時，帶寬是否會下降（下降明顯不符合預期）；

2. 更關注帶寬的峰值，每次算到的帶寬峰值，可以只關注 in 或者 out；

3. 平均值，在數據量遞增的情況下，可能無法體現最終的結果；

4. 請確保數據量足夠大，可以壓到帶寬上限（通過調整 b、e 或者 n 選項）。

常用參數及解釋

- GPU 數量

- -t,--nthreads 每個進程的線程數量配置， 默認 1；

- -g,--ngpus 每個線程的 GPU 數量，默認 1；

- 數據大小配置

- -b,--minbytes 開始的最小數據量，默認 32M；

- -e,--maxbytes 結束的最大數據量，默認 32M；

- 數據步長設置

- -i,--stepbytes 每次增加的數據量，默認: 1M；

- -f,--stepfactor 每次增加的倍數，默認禁用；

- NCCL 操作相關配置

- -o,--op 指定那種操作為reduce，僅適用于Allreduce、Reduce或ReduceScatter等縮減操作。默認值為：求和（Sum）；

- -d,--datatype 指定使用哪種數據類型，默認 : Float；

- 性能相關配置

- -n,--iters 每次操作（一次發送）循環多少次，默認 : 20；

- -w,--warmup_iters 預熱迭代次數（不計時），默認：5；

- -m,--agg_iters 每次迭代中要聚合在一起的操作數，默認：1；

- -a,--average <0/1/2/3> 在所有 ranks 計算均值作為最終結果 (MPI=1 only). <0=Rank0,1=Avg,2=Min,3=Max>，默認：1；

- 測試相關配置

- -p,--parallel_init <0/1> 使用線程并行初始化 NCCL，默認: 0；

- -c,--check <0/1> 檢查結果的正確性。在大量GPU上可能會非常慢，默認：1；

- -z,--blocking <0/1> 使NCCL集合阻塞，即在每個集合之后讓CPU等待和同步，默認：0；

- -G,--cudagraph 將迭代作為CUDA圖形捕獲，然后重復指定的次數，默認：0；

實驗測試

完成硬件、軟件的選型和配置后，下一步將進行實踐測試。

獲取LLaMA-Factory源碼包

因為網絡問題很難直接通過git clone命令行拉取，建議通過打包下載后自己上傳的方式進行：

noone@MacBook-Air Downloads % scp LLaMA-Factory-0.8.3.zip root@10.230.1.13:/tmp

[root@server3 AIGC]# pwd
/home/lichao/AIGC
[root@server3 AIGC]# cp /tmp/LLaMA-Factory-0.8.3.zip ./
[root@server3 AIGC]# unzip LLaMA-Factory-0.8.3.zip
[root@server3 AIGC]# cd LLaMA-Factory-0.8.3
[root@server3 LLaMA-Factory-0.8.3]# ll
總用量 128
drwxr-xr-x  2 root root    83 9月  13 05:04 assets
drwxr-xr-x  2 root root   122 9月   6 08:26 cache
-rw-r--r--  1 root root  1378 7月  18 19:36 CITATION.cff
drwxr-xr-x  6 root root  4096 9月  13 05:03 data
drwxr-xr-x  4 root root    43 7月  18 19:36 docker
drwxr-xr-x  5 root root    44 7月  18 19:36 evaluation
drwxr-xr-x 10 root root   182 7月  18 19:36 examples
-rw-r--r--  1 root root 11324 7月  18 19:36 LICENSE
-rw-r--r--  1 root root   242 7月  18 19:36 Makefile
-rw-r--r--  1 root root    33 7月  18 19:36 MANIFEST.in
-rw-r--r--  1 root root   645 7月  18 19:36 pyproject.toml
-rw-r--r--  1 root root 44424 7月  18 19:36 README.md
-rw-r--r--  1 root root 44093 7月  18 19:36 README_zh.md
-rw-r--r--  1 root root   245 7月  18 19:36 requirements.txt
drwxr-xr-x  3 root root    16 9月   6 18:48 saves
drwxr-xr-x  2 root root   219 7月  18 19:36 scripts
-rw-r--r--  1 root root  3361 7月  18 19:36 setup.py
drwxr-xr-x  4 root root   101 9月   6 08:22 src
drwxr-xr-x  5 root root    43 7月  18 19:36 tests
[root@server3 LLaMA-Factory-0.8.3]# 

安裝LLaMA-Factory，并進行驗證

[root@server3 LLaMA-Factory-0.8.3]# pip install -e ".[torch,metrics]"
[root@server3 LLaMA-Factory-0.8.3]# llamafactory-cli version
[2024-09-23 08:51:28,722] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
----------------------------------------------------------
| Welcome to LLaMA Factory, version 0.8.3                |
|                                                        |
| Project page: https://github.com/hiyouga/LLaMA-Factory |
----------------------------------------------------------
[root@server3 LLaMA-Factory-0.8.3]# 

下載訓練時所需的預訓練模型和數據集

根據當前GPU服務器所配置的GPU硬件規格，選擇適合的訓練方法、模型和數據集。

GPU型號：NVIDIA GeForce RTX 4060 Ti 16GB

預訓練模型：Qwen/Qwen1.5-0.5B-Chat

數據集：identity、alpaca_zh_demo

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install
git clone https://hf-mirror.com/Qwen/Qwen1.5-0.5B-Chat
# If you want to clone without large files - just their pointers
GIT_LFS_SKIP_SMUDGE=1 git clone https://hf-mirror.com/Qwen/Qwen1.5-0.5B-Chat

因為網絡問題通過命令行很難直接下載，這里使用huggingface的國內鏡像站拉取預訓練模型數據，并使用“GIT_LFS_SKIP_SMUDGE=1”變量跳過大文件，隨后手工下載后再上傳。

如果覺得麻煩，也可以安裝使用huggingface的命令行工具，下載預訓練模型和數據集。同樣地，安裝完成后，需要配置一些環境變量（使用鏡像站hf-mirror.com）來解決網絡問題。

1. 安裝依賴
[root@server3 LLaMA-Factory-0.8.3]# pip3 install -U huggingface_hub
2. 設置環境變量
[root@server3 LLaMA-Factory-0.8.3]# export HF_ENDPOINT=https://hf-mirror.com
可以寫入 ~/.bashrc 永久生效。
3. 確認環境變量生效
[root@server3 LLaMA-Factory-0.8.3]# huggingface-cli env

Copy-and-paste the text below in your GitHub issue.

- huggingface_hub version: 0.24.5
- Platform: Linux-3.10.0-1160.118.1.el7.x86_64-x86_64-with-glibc2.17
- Python version: 3.11.9
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Token path ?: /root/.cache/huggingface/token
- Has saved token ?: True
- Who am I ?: richard-open-source
- Configured git credential helpers: 
- FastAI: N/A
- Tensorflow: N/A
- Torch: 2.4.0
- Jinja2: 3.1.4
- Graphviz: N/A
- keras: N/A
- Pydot: N/A
- Pillow: 10.4.0
- hf_transfer: N/A
- gradio: 4.43.0
- tensorboard: N/A
- numpy: 1.26.4
- pydantic: 2.9.0
- aiohttp: 3.10.3
- ENDPOINT: https://hf-mirror.com
- HF_HUB_CACHE: /root/.cache/huggingface/hub
- HF_ASSETS_CACHE: /root/.cache/huggingface/assets
- HF_TOKEN_PATH: /root/.cache/huggingface/token
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False
- HF_HUB_ETAG_TIMEOUT: 10
- HF_HUB_DOWNLOAD_TIMEOUT: 10

[root@server3 LLaMA-Factory-0.8.3]# 
4.1 下載模型
[root@server3 LLaMA-Factory-0.8.3]# huggingface-cli download --resume-download Qwen/Qwen1.5-0.5B-Chat --local-dir ./models/Qwen1.5-0.5B-Chat
4.2 下載數據集
[root@server3 LLaMA-Factory-0.8.3]# huggingface-cli download --repo-type dataset --resume-download alpaca_zh_demo --local-dir ./datasets/alpaca_zh_demo

下載預訓練模型

[root@server3 AIGC]# mkdir models
[root@server3 AIGC]# cd models/
[root@server3 models]# GIT_LFS_SKIP_SMUDGE=1 git clone https://hf-mirror.com/Qwen/Qwen1.5-0.5B-Chat
[root@server3 models]# tree -h Qwen1.5-0.5B-Chat/
Qwen1.5-0.5B-Chat/
├── [ 656]  config.json
├── [ 661]  config.json.raw
├── [ 206]  generation_config.json
├── [7.1K]  LICENSE
├── [1.6M]  merges.txt
├── [1.2G]  model.safetensors
├── [4.2K]  README.md
├── [1.3K]  tokenizer_config.json
├── [6.7M]  tokenizer.json
└── [2.6M]  vocab.json

0 directories, 10 files
[root@server3 models]# 

下載數據集

默認情況下，LLaMA-Factory項目文件下的data目錄，自帶了一些本地數據集可直接使用。

[root@server3 LLaMA-Factory-0.8.3]# tree -h data/
data/
├── [841K]  alpaca_en_demo.json
├── [621K]  alpaca_zh_demo.json
├── [  32]  belle_multiturn
│   └── [2.7K]  belle_multiturn.py
├── [733K]  c4_demo.json
├── [ 13K]  dataset_info.json
├── [1.5M]  dpo_en_demo.json
├── [833K]  dpo_zh_demo.json
├── [722K]  glaive_toolcall_en_demo.json
├── [665K]  glaive_toolcall_zh_demo.json
├── [  27]  hh_rlhf_en
│   └── [3.3K]  hh_rlhf_en.py
├── [ 20K]  identity.json
├── [892K]  kto_en_demo.json
├── [  45]  mllm_demo_data
│   ├── [ 12K]  1.jpg
│   ├── [ 22K]  2.jpg
│   └── [ 16K]  3.jpg
├── [3.1K]  mllm_demo.json
├── [9.8K]  README.md
├── [9.2K]  README_zh.md
├── [  27]  ultra_chat
│   └── [2.3K]  ultra_chat.py
└── [1004K]  wiki_demo.txt

4 directories, 20 files
[root@server3 LLaMA-Factory-0.8.3]# 

使用準備好的模型與數據集，在單機上進行訓練測試

LLaMA-Factory支持通過WebUI微調大語言模型。在完成安裝后，我們可以使用WebUI進行快速調測驗證，沒問題后可使用命令行工具進行多機分布式訓練。

[root@server3 LLaMA-Factory-0.8.3]# llamafactory-cli webui
[2024-09-23 17:54:45,786] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
Running on local URL:  http://0.0.0.0:7861

To create a public link, set `share=True` in `launch()`.

使用命令行運行多機分布式訓練任務

1. 準備目錄
[root@server3 LLaMA-Factory-0.8.3]# mkdir asterun
[root@server3 LLaMA-Factory-0.8.3]# mkdir -p asterun/saves/qwen/full/sft
2. 根據集群環境和訓練任務，準備分布式訓練的配置文件
[root@server3 LLaMA-Factory-0.8.3]# cat asterun/qwen_full_sft_ds2.yaml 
### model
model_name_or_path: /home/lichao/AIGC/models/Qwen1.5-0.5B-Chat

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z2_config.json

### dataset
dataset: identity,alpaca_zh_demo
template: llama3
cutoff_len: 1024
max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: asterun/saves/qwen/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

report_to: tensorboard
logging_dir: asterun/saves/qwen/full/sft/runs

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500
[root@server3 LLaMA-Factory-0.8.3]# 
3. 用同樣的方式，在Server1和Server2上準備運行環境
步驟略。
4. 依次在集群中的3個GPU節點上啟動分布式訓練任務
主節點rank0：
[root@server3 LLaMA-Factory-0.8.3]# FORCE_TORCHRUN=1 NNODES=3 RANK=0 MASTER_ADDR=172.16.0.13 MASTER_PORT=29500 llamafactory-cli train asterun/qwen_full_sft_ds2.yaml
從節點rank1：
[root@server2 LLaMA-Factory-0.8.3]# FORCE_TORCHRUN=1 NNODES=3 RANK=1 MASTER_ADDR=172.16.0.13 MASTER_PORT=29500 llamafactory-cli train asterun/qwen_full_sft_ds2.yaml
從節點rank2：
[root@server1 LLaMA-Factory-0.8.3]# FORCE_TORCHRUN=1 NNODES=3 RANK=2 MASTER_ADDR=172.16.0.13 MASTER_PORT=29500 llamafactory-cli train asterun/qwen_full_sft_ds2.yaml

推理測試

安裝GGUF庫

下載llama.cpp源碼包到服務器，解壓到工作目錄
[root@server3 AIGC]# unzip llama.cpp-master.zip
[root@server3 AIGC]# cd llama.cpp-master
[root@server3 llama.cpp-master]# ll
總用量 576
-rw-r--r--  1 root root  33717 9月  26 11:38 AUTHORS
drwxr-xr-x  2 root root     37 9月  26 11:38 ci
drwxr-xr-x  2 root root    164 9月  26 11:38 cmake
-rw-r--r--  1 root root   6591 9月  26 11:38 CMakeLists.txt
-rw-r--r--  1 root root   3164 9月  26 11:38 CMakePresets.json
drwxr-xr-x  3 root root   4096 9月  26 11:38 common
-rw-r--r--  1 root root   2256 9月  26 11:38 CONTRIBUTING.md
-rwxr-xr-x  1 root root 199470 9月  26 11:38 convert_hf_to_gguf.py
-rwxr-xr-x  1 root root  15993 9月  26 11:38 convert_hf_to_gguf_update.py
-rwxr-xr-x  1 root root  19106 9月  26 11:38 convert_llama_ggml_to_gguf.py
-rwxr-xr-x  1 root root  14901 9月  26 11:38 convert_lora_to_gguf.py
drwxr-xr-x  4 root root    109 9月  26 11:38 docs
drwxr-xr-x 43 root root   4096 9月  26 11:38 examples
-rw-r--r--  1 root root   1556 9月  26 11:38 flake.lock
-rw-r--r--  1 root root   7469 9月  26 11:38 flake.nix
drwxr-xr-x  5 root root     85 9月  26 11:38 ggml
drwxr-xr-x  6 root root    116 9月  26 11:38 gguf-py
drwxr-xr-x  2 root root    154 9月  26 11:38 grammars
drwxr-xr-x  2 root root     21 9月  26 11:38 include
-rw-r--r--  1 root root   1078 9月  26 11:38 LICENSE
-rw-r--r--  1 root root  50865 9月  26 11:38 Makefile
drwxr-xr-x  2 root root    163 9月  26 11:38 media
drwxr-xr-x  2 root root   4096 9月  26 11:38 models
-rw-r--r--  1 root root    163 9月  26 11:38 mypy.ini
-rw-r--r--  1 root root   2044 9月  26 11:38 Package.swift
drwxr-xr-x  3 root root     40 9月  26 11:38 pocs
-rw-r--r--  1 root root 124786 9月  26 11:38 poetry.lock
drwxr-xr-x  2 root root   4096 9月  26 11:38 prompts
-rw-r--r--  1 root root   1280 9月  26 11:38 pyproject.toml
-rw-r--r--  1 root root    528 9月  26 11:38 pyrightconfig.json
-rw-r--r--  1 root root  28481 9月  26 11:38 README.md
drwxr-xr-x  2 root root   4096 9月  26 11:38 requirements
-rw-r--r--  1 root root    505 9月  26 11:38 requirements.txt
drwxr-xr-x  2 root root   4096 9月  26 11:38 scripts
-rw-r--r--  1 root root   5090 9月  26 11:38 SECURITY.md
drwxr-xr-x  2 root root     97 9月  26 11:38 spm-headers
drwxr-xr-x  2 root root    289 9月  26 11:38 src
drwxr-xr-x  2 root root   4096 9月  26 11:38 tests
[root@server3 llama.cpp-master]# 

進入gguf-py子目錄，安裝GGUF庫
[root@server3 llama.cpp-master]# cd gguf-py
[root@server3 gguf-py]# ll
總用量 12
drwxr-xr-x 2 root root   40 9月  26 11:38 examples
drwxr-xr-x 2 root root  230 9月  26 11:38 gguf
-rw-r--r-- 1 root root 1072 9月  26 11:38 LICENSE
-rw-r--r-- 1 root root 1049 9月  26 11:38 pyproject.toml
-rw-r--r-- 1 root root 2719 9月  26 11:38 README.md
drwxr-xr-x 2 root root  151 9月  26 11:38 scripts
drwxr-xr-x 2 root root   71 9月  26 11:38 tests
[root@server3 gguf-py]# pip install --editable .
Looking in indexes: https://mirrors.aliyun.com/pypi/simple/
Obtaining file:///home/lichao/AIGC/llama.cpp-master/gguf-py
  Installing build dependencies ... done
  Checking if build backend supports build_editable ... done
  Getting requirements to build editable ... done
  Preparing editable metadata (pyproject.toml) ... done
Requirement already satisfied: numpy>=1.17 in /home/lichao/opt/python3.11.9/lib/python3.11/site-packages (from gguf==0.10.0) (1.26.4)
Requirement already satisfied: pyyaml>=5.1 in /home/lichao/opt/python3.11.9/lib/python3.11/site-packages (from gguf==0.10.0) (6.0.2)
Requirement already satisfied: sentencepiece<=0.2.0,?>=0.1.98 in /home/lichao/opt/python3.11.9/lib/python3.11/site-packages (from gguf==0.10.0) (0.2.0)
Requirement already satisfied: tqdm>=4.27 in /home/lichao/opt/python3.11.9/lib/python3.11/site-packages (from gguf==0.10.0) (4.66.5)
Building wheels for collected packages: gguf
  Building editable for gguf (pyproject.toml) ... done
  Created wheel for gguf: filename=gguf-0.10.0-py3-none-any.whl size=3403 sha256=4a0851426e263076c64c9854be9dfe95493844062484d001fddb08c1be5fa2ca
  Stored in directory: /tmp/pip-ephem-wheel-cache-iiq8ofh3/wheels/80/80/9b/c6c23d750f4bd20fc0c2c75e51253d89c61a2369247fb694db
Successfully built gguf
Installing collected packages: gguf
Successfully installed gguf-0.10.0
WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable.It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning.
[root@server3 gguf-py]# 

模型格式轉換

將之前微調訓練生成的safetensors格式的模型，轉換為gguf格式
[root@server3 gguf-py]# cd .. 
[root@server3 llama.cpp-master]# python3 convert_hf_to_gguf.py /home/lichao/AIGC/LLaMA-Factory-0.8.3/asterun/saves/qwen/full/sft
INFO:hf-to-gguf:Loading model: sft
INFO:gguf.gguf_writer:gguf: This GGUF file is for Little Endian only
INFO:hf-to-gguf:Exporting model...
INFO:hf-to-gguf:gguf: loading model part 'model.safetensors'
INFO:hf-to-gguf:output.weight,             torch.bfloat16 --> F16, shape = {1024, 151936}
INFO:hf-to-gguf:token_embd.weight,         torch.bfloat16 --> F16, shape = {1024, 151936}
INFO:hf-to-gguf:blk.0.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.0.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.0.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.0.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.0.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.0.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.0.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.0.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.0.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.0.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.0.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.0.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.1.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.1.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.1.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.1.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.1.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.1.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.1.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.1.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.1.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.1.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.1.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.1.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.10.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.10.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.10.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.10.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.10.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.10.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.10.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.10.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.10.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.10.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.10.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.10.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.11.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.11.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.11.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.11.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.11.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.11.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.11.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.11.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.11.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.11.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.11.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.11.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.12.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.12.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.12.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.12.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.12.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.12.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.12.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.12.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.12.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.12.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.12.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.12.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.13.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.13.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.13.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.13.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.13.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.13.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.13.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.13.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.13.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.13.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.13.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.13.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.14.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.14.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.14.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.14.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.14.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.14.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.14.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.14.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.14.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.14.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.14.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.14.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.15.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.15.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.15.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.15.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.15.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.15.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.15.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.15.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.15.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.15.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.15.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.15.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.16.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.16.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.16.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.16.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.16.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.16.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.16.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.16.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.16.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.16.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.16.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.16.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.17.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.17.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.17.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.17.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.17.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.17.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.17.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.17.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.17.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.17.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.17.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.17.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.18.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.18.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.18.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.18.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.18.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.18.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.18.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.18.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.18.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.18.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.18.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.18.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.19.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.19.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.19.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.19.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.19.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.19.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.19.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.19.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.19.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.19.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.19.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.19.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.2.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.2.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.2.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.2.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.2.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.2.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.2.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.2.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.2.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.2.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.2.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.2.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.20.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.20.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.20.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.20.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.20.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.20.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.20.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.20.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.20.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.20.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.20.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.20.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.21.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.21.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.21.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.21.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.21.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.21.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.21.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.21.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.21.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.21.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.21.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.21.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.22.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.22.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.22.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.22.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.22.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.22.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.22.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.22.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.22.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.22.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.22.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.22.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.23.attn_norm.weight,   torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.23.ffn_down.weight,    torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.23.ffn_gate.weight,    torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.23.ffn_up.weight,      torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.23.ffn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.23.attn_k.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.23.attn_k.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.23.attn_output.weight, torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.23.attn_q.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.23.attn_q.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.23.attn_v.bias,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.23.attn_v.weight,      torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.3.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.3.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.3.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.3.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.3.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.3.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.3.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.3.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.3.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.3.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.3.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.3.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.4.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.4.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.4.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.4.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.4.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.4.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.4.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.4.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.4.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.4.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.4.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.4.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.5.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.5.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.5.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.5.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.5.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.5.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.5.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.5.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.5.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.5.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.5.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.5.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.6.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.6.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.6.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.6.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.6.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.6.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.6.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.6.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.6.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.6.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.6.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.6.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.7.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.7.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.7.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.7.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.7.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.7.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.7.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.7.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.7.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.7.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.7.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.7.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.8.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.8.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.8.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.8.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.8.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.8.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.8.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.8.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.8.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.8.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.8.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.8.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.9.attn_norm.weight,    torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.9.ffn_down.weight,     torch.bfloat16 --> F16, shape = {2816, 1024}
INFO:hf-to-gguf:blk.9.ffn_gate.weight,     torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.9.ffn_up.weight,       torch.bfloat16 --> F16, shape = {1024, 2816}
INFO:hf-to-gguf:blk.9.ffn_norm.weight,     torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.9.attn_k.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.9.attn_k.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.9.attn_output.weight,  torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.9.attn_q.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.9.attn_q.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:blk.9.attn_v.bias,         torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:blk.9.attn_v.weight,       torch.bfloat16 --> F16, shape = {1024, 1024}
INFO:hf-to-gguf:output_norm.weight,        torch.bfloat16 --> F32, shape = {1024}
INFO:hf-to-gguf:Set meta model
INFO:hf-to-gguf:Set model parameters
INFO:hf-to-gguf:gguf: context length = 32768
INFO:hf-to-gguf:gguf: embedding length = 1024
INFO:hf-to-gguf:gguf: feed forward length = 2816
INFO:hf-to-gguf:gguf: head count = 16
INFO:hf-to-gguf:gguf: key-value head count = 16
INFO:hf-to-gguf:gguf: rope theta = 1000000.0
INFO:hf-to-gguf:gguf: rms norm epsilon = 1e-06
INFO:hf-to-gguf:gguf: file type = 1
INFO:hf-to-gguf:Set model tokenizer
INFO:gguf.vocab:Adding 151387 merge(s).
INFO:gguf.vocab:Setting special token type eos to 151646
INFO:gguf.vocab:Setting special token type pad to 151643
INFO:gguf.vocab:Setting special token type bos to 151643
INFO:gguf.vocab:Setting chat_template to {% if messages[0]['role'] == 'system' %}{% set system_message = messages[0]['content'] %}{% endif %}{% if system_message is defined %}{{ '<|start_header_id|?>system<|end_header_id|?>

' + system_message + '<|eot_id|?>' }}{% endif %}{% for message in messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|start_header_id|?>user<|end_header_id|?>

' + content + '<|eot_id|?><|start_header_id|?>assistant<|end_header_id|?>

' }}{% elif message['role'] == 'assistant' %}{{ content + '<|eot_id|?>' }}{% endif %}{% endfor %}
INFO:hf-to-gguf:Set model quantization version
INFO:gguf.gguf_writer:Writing the following files:
INFO:gguf.gguf_writer:/home/lichao/AIGC/LLaMA-Factory-0.8.3/asterun/saves/qwen/full/sft/Sft-620M-F16.gguf: n_tensors = 291, total_size = 1.2G
Writing: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1.24G/1.24G [00:03

安裝Ollama

OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http://127.0.0.1:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:5m0s OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/root/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:0 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: ROCR_VISIBLE_DEVICES: http_proxy: https_proxy: no_proxy:]"
time=2024-09-26T12:04:20.753+02:00 level=INFO source=images.go:753 msg="total blobs: 0"
time=2024-09-26T12:04:20.754+02:00 level=INFO source=images.go:760 msg="total unused blobs removed: 0"
time=2024-09-26T12:04:20.754+02:00 level=INFO source=routes.go:1200 msg="Listening on 127.0.0.1:11434 (version 0.3.12)"
time=2024-09-26T12:04:20.755+02:00 level=INFO source=common.go:135 msg="extracting embedded files" dir=/tmp/ollama316805737/runners
time=2024-09-26T12:04:39.145+02:00 level=INFO source=common.go:49 msg="Dynamic LLM libraries" runners="[cpu cpu_avx cpu_avx2 cuda_v11 cuda_v12 rocm_v60102]"
time=2024-09-26T12:04:39.145+02:00 level=INFO source=gpu.go:199 msg="looking for compatible GPUs"
time=2024-09-26T12:04:39.283+02:00 level=INFO source=types.go:107 msg="inference compute" id=GPU-2d337ad0-020d-0464-2d00-715b0d00c7ba library=cuda variant=v12 compute=8.9 driver=12.4 name="NVIDIA GeForce RTX 4060 Ti" total="15.7 GiB" available="15.6 GiB"

注冊模型

打開一個新的terminal

[root@server3 AIGC]# cd LLaMA-Factory-0.8.3/asterun/
[root@server3 asterun]# ll
總用量 4
-rw-r--r-- 1 root root 817 9月  19 09:33 qwen_full_sft_ds2.yaml
drwxr-xr-x 3 root root  18 9月  13 10:28 saves
創建模型的Modelfile文件
[root@server3 asterun]# touch qwen_full_sft_ds2.ollama.Modelfile
[root@server3 asterun]# vim qwen_full_sft_ds2.ollama.Modelfile 
[root@server3 asterun]# cat qwen_full_sft_ds2.ollama.Modelfile 
FROM /home/lichao/AIGC/LLaMA-Factory-0.8.3/asterun/saves/qwen/full/sft/qwen-sft-620M-F16.gguf
[root@server3 asterun]# cd ../..
使用Modelfile注冊模型
[root@server3 AIGC]# ollama create qwen-full-sft -f ./LLaMA-Factory-0.8.3/asterun/qwen_full_sft_ds2.ollama.Modelfile 
transferring model data 100% 
using existing layer sha256:19d794be57081c1a5aa7e03c4045a0fdc5b8a40f080f0c550ab38033cf0d5d58 
creating new layer sha256:c33681b055686143e7d6e0bb0f1054c9910c05c3f4ab16932fbc567a8961929a 
writing manifest 
success 
[root@server3 AIGC]# 

使用注冊好的模型運行推理服務
[root@server3 AIGC]# ollama run qwen-full-sft
>>> who are you?
<|im_end|?>
我是 {{name}}，一個由 {{author}} 開發的人工智能助手，我可以幫助用戶查詢信息、安排日程、提供建議等。

>>> can you speak english?
 I am an AI assistant developed by {{author}}.

>>> 好吧，用中文交流吧。
沒問題。

>>> 你喜歡中國哪個城市？
每個城市都有其獨特的魅力，各具特色，比如：
成都：美食之都，生活悠閑。
北京：歷史悠久，文化豐富。
杭州：風景優美，以西湖聞名。
上海：現代化大都市，經濟繁榮。

>>> 感謝，再見
好的，我是個人工智能助手，很高興見到您。

>>> exit
[root@server3 AIGC]# 

至此，已完成分布式計算環境的搭建與測試

審核編輯 黃宇?