search

Oneflow-Inc / libai

LiBai(李白): A Toolbox for Large-Scale Distributed Parallel Training
https://libai.readthedocs.io
Apache License 2.0
391 stars 55 forks source link

Qwen #537

Closed xiezipeng-ML closed 1 month ago

ShawnXuan commented 1 month ago

推理

python projects/Qwen/pipeline.py --model_path=/root/models/Qwen1.5-7B-Chat --mode=huggingface
python projects/Qwen/pipeline.py --model_path=/data0/hf_models/qwen2/Qwen1.5-7B-Chat --mode=huggingface --device=npu
python projects/Qwen/pipeline.py --model_path=/root/models/Qwen1.5-7B-Chat --mode=huggingface --device=xpu

训练

python projects/Qwen/utils/data_prepare.py
export NUM_GPUS=8
python3 -m oneflow.distributed.launch \
    --nproc_per_node ${NUM_GPUS} \
    --nnodes 1 \
    --node_rank 0 \
    --master_addr 127.0.0.1 \
    --master_port 12345 \
        tools/train_net.py --config-file=projects/Qwen/configs/qwen_sft.py \
            graph.enabled=True \
            train.input_placement_device="cuda" \
            train.dist.device_type="cuda" \
            train.dist.pipeline_parallel_size=${NUM_GPUS}

A100-PCIE-40GB x 4 OOM

export NUM_GPUS=1
python3 -m oneflow.distributed.launch \
    --nproc_per_node ${NUM_GPUS} \
    --nnodes 1 \
    --node_rank 0 \
    --master_addr 127.0.0.1 \
    --master_port 12345 \
        tools/train_net.py --config-file=projects/Qwen/configs/qwen_sft.py \
            graph.enabled=False \
            train.input_placement_device="xpu" \
            train.dist.device_type="xpu" \
            train.dist.pipeline_parallel_size=${NUM_GPUS}