ms-swift/train/keye_timing_test.sh

#!/bin/bash
# 专门用于单条数据时间测试的训练脚本
PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True'
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export NPROC_PER_NODE=8 
export VIDEO_MAX_PIXELS=50176
export IMAGE_MAX_PIXELS=50176
export MAX_PIXELS=50176

# 明确设置分布式训练的设备
export MASTER_ADDR=localhost
export MASTER_PORT=29500
export WORLD_SIZE=8
export RANK=0

export SEQUENCE_PARALLEL_IMPL=ulysses
# 根据序列长度调整CELOSS_PARALLEL_SIZE，65536/4=16384
export CELOSS_PARALLEL_SIZE=16384 

swift sft \
    --model Qwen/Qwen2.5-VL-7B-Instruct \
    --train_type full \
    --deepspeed zero3 \
    --attn_impl flash_attn \
    --sequence_parallel_size 4 \
    --freeze_vit false \
    --freeze_aligner false \
    --freeze_llm false \
    --use_hf true \
    --dataset 'datasets/jsonl/filter_llava_350_550#32' \
    --dataset_num_proc 4 \
    --split_dataset_ratio 0.001 \
    --save_strategy steps \
    --save_steps 1 \
    --torch_dtype bfloat16 \
    --max_steps 1 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --learning_rate 5e-6 \
    --gradient_accumulation_steps 1 \
    --eval_steps 1000 \
    --save_total_limit 5 \
    --logging_steps 1 \
    --max_length 76000 \
    --warmup_ratio 0.05 \
    --dataloader_num_workers 1 \
    --gradient_checkpointing true \
    --max_grad_norm 1.0 \
    --use_liger_kernel true \
    --loss_scale default \
    --load_from_cache_file true \
    --save_safetensors true \
    --report_to tensorboard \
    --logging_dir output/runs \
    --output_dir output/timing_test_single_sample