[Usage] Deepspeed MoE hangs when EP_SIZE > 1

Describe the issue

Issue: When I try EP_SIZE > 1 (like EP_SIZE = 2), the program hangs at beginning. It hasn't even finished the first batch yet. May I ask if you have encountered this issue before?

Environment:

GPU: 8×A100-80G or others? 8×A100-80G 
Deepspeed version: 0.9.5
Torch version: 2.0.1
Transformers version: 4.37.0
Tokenizers version: 0.15.1

Command:

#!/bin/bash

moe_mode="sparse"
num_experts=4
top_k_experts=2
use_residual=False
router_aux_loss_coef=0.01
JSON_FOLDER="ft_json"
IMAGE_FOLDER="train_image_video"
cd ~/jason/MoE-LLaVA

HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 deepspeed moellava/train/train_mem.py \
    --moe_enable True --num_experts ${num_experts} --top_k_experts ${top_k_experts} --capacity_factor 1.5 \
    --moe_mode ${moe_mode} --use_residual ${use_residual} --router_aux_loss_coef ${router_aux_loss_coef} \
    --train_modules fc1 fc2 wg \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./checkpoints/MoE-LLaVA-Phi2-2.7B-4e  \
    --version phi \
    --data_path ${JSON_FOLDER}/llava_image_tune_.json ${JSON_FOLDER}/nlp_tune.json \
    --image_folder ${IMAGE_FOLDER} \
    --image_tower openai/clip-vit-large-patch14-336 \
    --image_projector_type mlp2x_gelu \
    --mm_vision_select_layer -2 \
    --mm_use_im_start_end False \
    --mm_use_im_patch_token False \
    --image_aspect_ratio pad \
    --group_by_modality_length True \
    --bf16 True \
    --output_dir ./checkpoints/llavaphi-2.7b-finetune-moe \
    --num_train_epochs 1 \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 4 \
    --gradient_accumulation_steps 2 \
    --evaluation_strategy "no" \
    --save_strategy "steps" \
    --save_steps 24000 \
    --save_total_limit 1 \
    --learning_rate 2e-5 \
    --weight_decay 0. \
    --warmup_ratio 0.03 \
    --lr_scheduler_type "cosine" \
    --logging_steps 1 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 4 \
    --lazy_preprocess True \
    --report_to tensorboard \
    --cache_dir "./cache_dir"

Log:

No existing process group found, creating a new group named: ep_size_2
No existing process group found, creating a new group named: ep_size_2
Rank: 2 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 6 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 0 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 3 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 5 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 7 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 1 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
Rank: 4 partition count [8, 8, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4] and sizes[(106024960, False), (26240, False), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (39321600, True), (26214400, True), (102400, True)] 
  0%|                                                                                                                                | 0/1232 [00:00<?, ?it/s]

Screenshots: You may attach screenshots if it better explains the issue.

PKU-YuanGroup / MoE-LLaVA

[Usage] Deepspeed MoE hangs when EP_SIZE > 1 #65

Describe the issue