README.md

---
license: apache-2.0
datasets:
- Vi-VLM/Vista
language:
- vi
---

- Deploying Gradio: Going to MoE-LLaVA serving tutorial and follow, or quickly, can use 👇👇👇
```bash
deepspeed --include localhost:0 moellava/serve/gradio_web_server.py \
    --model-path="tuanio/moe-llava-qwen1.5-0.5b-vista_reasoning-2ep"
```

Or CLI: 
```bash
deepspeed --include localhost:0 moellava/serve/cli.py \
    --model-path "tuanio/moe-llava-qwen1.5-0.5b-vista_reasoning-2ep" \
    --image-file "data/llm_data/coco2017/train2017/000000391895.jpg"
```

---

- Training script:
```bash
moe_mode="sparse"
num_experts=4
top_k_experts=2
use_residual=False
router_aux_loss_coef=0.01

ROOT_DATA=data/llm_data

WANDB_PROJECT=chart-vision-llm CUDA_VISIBLE_DEVICES=0,1,2,3,4 deepspeed --include localhost:2,3,4 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.gate_proj mlp.up_proj mlp.down_proj wg \
    --deepspeed ./scripts/zero2.json \
    --model_name_or_path ./checkpoints/ft-llava-qwen1.5-0.5b-vista_llava-merged-2ep \
    --version qwen \
    --data_path $ROOT_DATA/json_files/vista_reasoning.json \
    --image_folder $ROOT_DATA/coco2017/train2017 \
    --image_tower google/siglip-base-patch16-256-multilingual \
    --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/ft-moe-llava-qwen1.5-0.5b-vista_reasoning-2ep \
    --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 50 \
    --tf32 True \
    --model_max_length 2048 \
    --gradient_checkpointing True \
    --dataloader_num_workers 8 \
    --lazy_preprocess True \
    --report_to wandb \
    --cache_dir "./cache_dir" \
    --run_name ft-moe-llava-qwen1.5-0.5b-vista_reasoning-2ep
```