Moe finetune error

[AdonLee072348]

The following error occurred while running the script finetune_moe.sh:
The model has moe layers, but None of the param groups are marked as MoE. Create a param group with 'moe' key set to True before creating optimizer.

[LinB203]

Could you post your command & transformers version & deepspeed version & torch version. Thanks

[AdonLee072348]

Thanks for your reply. The script is MoE-LLaVA/scripts/v1_5/qwen/finetune_moe.sh

torch version: 2.0.1+cu117
deepspeed version: 0.13.1
transformers version: 4.33.0

command:
moe_mode="sparse"
num_experts=4
top_k_experts=2
use_residual=False
router_aux_loss_coef=0.01
JSON_FOLDER="./MoE-LLaVA/train_json"
IMAGE_FOLDER="./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}
--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/llava-moe-qwen-1.8b-pretrain-finetune
--version qwen
--data_path ${JSON_FOLDER}/llava_image_tune_.json ${JSON_FOLDER}/nlp_tune.json
--image_folder ${IMAGE_FOLDER}
--image_tower ./pretrain_weight/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/llava-moe-qwen-1.8b-pretrain-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"

[whalefa1I]

Could you post your command & transformers version & deepspeed version & torch version. Thanks

pip list|grep -E "torch|transformers|accelerate|deepspeed"
accelerate                    0.23.0
deepspeed                     0.9.5+cf42c28        
open-clip-torch               2.20.0
pytorch-lightning             1.7.7
pytorch-metric-learning       2.3.0
pytorch-wavelets              1.3.0
pytorch-wpe                   0.0.1
pytorch3d                     0.7.4
rotary-embedding-torch        0.3.0
taming-transformers-rom1504   0.0.6
torch                         2.0.1+cu118
torch-complex                 0.4.3
torch-scatter                 2.1.1
torchaudio                    2.0.2+cu118
torchmetrics                  0.11.4
torchsummary                  1.5.1
torchvision                   0.15.2+cu118
transformers                  4.34.1
transformers-stream-generator 0.0.4

command

HF_DATASETS_OFFLINE=1 TRANSFORMERS_OFFLINE=1 deepspeed moellava/train/train_mem.py     --moe_enable True --num_experts 4 --top_k_experts 2     --moe_mode sparse --use_residual False --router_aux_loss_coef 0.01     --train_modules fc1 fc2 wg     --deepspeed ./scripts/zero2.json     --model_name_or_path ./checkpoints/llavaqwen-1.8b-finetune     --version qwen     --data_path ./IMAGE_FOLDER/train_json/llava_image_tune_.json ./IMAGE_FOLDER/train_json/nlp_tune.json     --image_folder ./IMAGE_FOLDER     --image_tower ./cache/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/llavaqwen-1.8b-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"

[LinB203]

Have you try it on transformers==4.36.2 & deepspeed==0.9.5? We will check MoE-Qwen latter. And could you guys try Phi or StableLM？

[AdonLee072348]

Have you try it on transformers==4.36.2 & deepspeed==0.9.5? We will check MoE-Qwen latter. And could you guys try Phi or StableLM？

this error still occurred on transformers==4.36.2 & deepspeed==0.9.5

[LinB203]

Have you try it on transformers==4.36.2 & deepspeed==0.9.5? We will check MoE-Qwen latter. And could you guys try Phi or StableLM？

this error still occurred on transformers==4.36.2 & deepspeed==0.9.5

Copy that. I will check it as soon as possible.

[LinB203]

Hi, everyone. Sorry for that, we updated the new runing command to fix it. Checking out here

[whalefa1I]

Hi, everyone. Sorry for that, we updated the new runing command to fix it. Checking out here

Works for me, thx~!

[LinB203]

For training stage 3, we print the loss for each step to observe the router banlanced loss. If someone do not like it, just comment out this line.

[hxhcreate]

Hi, everyone. Sorry for that, we updated the new runing command to fix it. Checking out here

I still have the following error

AssertionError: The model has moe layers, but None of the param groups are marked as MoE. Create a param group with 'moe' key set to True before creating optimizer

here is my command

torchrun $DISTRIBUTED_ARGS 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 mlp.w1 mlp.w2 mlp.c_proj wg \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./checkpoints/llavaqwen1.5-1.8b-finetune \
    --version qwen \
    --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/llavaqwen-1.8b-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"

here is my package version

accelerate                0.21.0
deepspeed                 0.9.5
torch                     2.0.1
torchvision               0.15.2
transformers              4.37.0

[PangziZhang523]

I still have the following error !!!!!

AssertionError: The model has moe layers, but None of the param groups are marked as MoE. Create a param group with 'moe' key set to True before creating optimizer

here is my package version:
accelerate 0.21.0
deepspeed 0.9.5
torch 2.0.1
torchvision 0.15.2
transformers 4.37.0

[Wuyingwen]

Hi, everyone. Sorry for that, we updated the new runing command to fix it. Checking out here

I still have the following error

AssertionError: The model has moe layers, but None of the param groups are marked as MoE. Create a param group with 'moe' key set to True before creating optimizer

here is my command

torchrun $DISTRIBUTED_ARGS 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 mlp.w1 mlp.w2 mlp.c_proj wg \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./checkpoints/llavaqwen1.5-1.8b-finetune \
    --version qwen \
    --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/llavaqwen-1.8b-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"

here is my package version

accelerate                0.21.0
deepspeed                 0.9.5
torch                     2.0.1
torchvision               0.15.2
transformers              4.37.0

hi, I have the same environment setting as you. And meet the same problem when training a moe version of llama_llava. Have you solve the problem now?