# verl-main使用指南

## github链接

```
https://github.com/volcengine/verl
```

### mixtral_8x7B lora微调

#### 预训练权重

```shell
# 预训练权重开源地址
https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1

# bw地址
/public/opendas/DL_DATA/llm-models/Mixtral-8x7B-Instruct-v0.1
```

#### 数据集

```shell
# 开源地址
https://huggingface.co/datasets/openai/gsm8k/tree/main
```

#### 环境变量

```shell
# 具体根据verl-main目录修改
export PYTHONPATH=/public/home/fugx1/ds/verl-main:$PYTHONPATH
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
```

#### 拉起命令

```shell
# 具体根据verl-main目录修改
cd /public/home/fugx1/ds/verl-main/examples/sft/gsm8k

# 执行命令
bash run_mixtral_8x7B.sh 8 /public/home/fugx1/ds/verl-main/examples/sft/gsm8k
```

#### 脚本内容

```shell
# 需要根据实际路径修改
# data.train_files
# data.val_files
# model.partial_pretrain

set -x

if [ "$#" -lt 2 ]; then
    echo "Usage: run_mixtral_8x7B.sh <nproc_per_node> <save_path> [other_configs...]"
    exit 1
fi

nproc_per_node=$1
save_path=$2

source /opt/dtk/env.sh

# Shift the arguments so $@ refers to the rest
shift 2

torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
     -m verl.trainer.fsdp_sft_trainer \
    data.train_files=/public/home/fugx1/ds/gsm8k/gsm8k-train.parquet \
    data.val_files=/public/home/fugx1/ds/gsm8k/gsm8k-test.parquet \
    data.prompt_key='question' \
    data.response_key='answer' \
    +data.prompt_dict_keys=['question'] \
    +data.response_dict_keys=['answer'] \
    data.micro_batch_size_per_gpu=4 \
    model.partial_pretrain=/public/opendas/DL_DATA/llm-models/Mixtral-8x7B-Instruct-v0.1 \
    trainer.default_local_dir=$save_path \
    trainer.project_name=gsm8k-sft \
    trainer.experiment_name=gsm8k-sft-mixtral-8x7B-Instruct-v0.1 \
    trainer.total_epochs=1 \
    trainer.logger=['console'] \
    trainer.default_hdfs_dir=null $@

```

#### 依赖

```shell
# 有一些依赖在requirements.txt没有给出,需要在执行时报错后进行安装
vllm
pip install tensordict-0.6.2-cp310-cp310-manylinux1_x86_64.whl --no-deps
pip install orjson --no-deps
pip install cloudpickle --no-deps
pip install ray --no-deps
pip install msgpack --no-deps
pip install google --no-deps
pip install protobuf --no-deps
pip install jsonschema --no-deps
pip install referencing --no-deps
pip install rpds-py --no-deps
pip install hydra-core --no-deps
pip install codetiming --no-deps
pip install jsonschema_specifications --no-deps
pip install pytest --no-deps
pip install pluggy --no-deps
pip install exceptiongroup --no-deps
pip install iniconfig --no-deps
pip install omegaconf --no-deps
pip install antlr4 --no-deps
pip install antlr4-python3-runtime==4.9.3 --no-deps
pip install click --no-deps
pip install wandb --no-deps
pip install pydantic --no-deps
pip install annotated_types --no-deps
pip install msgspec --no-deps
pip install zmq --no-deps
pip install pyzmq --no-deps
pip install blake3 --no-deps
pip install cpuinfo --no-deps
pip install py-cpuinfo --no-deps
pip install openai --no-deps
pip install httpx --no-deps
pip install sniffio --no-deps
pip install anyio --no-deps
pip install distro --no-deps
pip install jiter --no-deps
pip install gguf --no-deps
pip install numa --no-deps

```

### DeepSeek V3 lora减层微调

#### 预训练权重

```shell
# huggingface链接
https://huggingface.co/deepseek-ai/DeepSeek-V3

# bw千卡集群目录
/public/opendas/DL_DATA/llm-models/DeepSeek-V3-bf16
```

#### 数据集

```shell
# 开源地址
https://huggingface.co/datasets/openai/gsm8k/tree/main
```

#### 环境变量

```shell
# 具体根据verl-main目录修改
export PYTHONPATH=/public/home/fugx1/ds/verl-main:$PYTHONPATH
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
```

#### 文件修改

```shell
1.modeling_deepseek.py
修改前:https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/modeling_deepseek.py
修改后:/public/opendas/DL_DATA/llm-models/DeepSeek-V3-bf16/modeling_deepseek.py

说明:DeepSeek-V3自带的modeling_deepseek.py代码不完整,有一个团队重新实现了包含DeepSeek-V3/R1 训练逻辑的 modeling 文件,这里对modeling_deepseek.py做了补全和修改。
参考链接:https://github.com/ScienceOne-AI/DeepSeek-671B-SFT-Guide/blob/main/model/DeepSeek-V3-BF16/modeling_deepseek.py


2.sft_trainer.yaml
位置:verl-main/verl/trainer/config/sft_trainer.yaml
修改点:
train_batch_size: 256 -> 128
min_num_params:  0 -> 1
trust_remote_code: False -> True

3.config.json
位置:/public/opendas/DL_DATA/llm-models/DeepSeek-V3-bf16/config.json
修改点:
num_hidden_layers 61 -> 4(8)
说明:单机最多放下8层,但是性能很差,这里先设置4层

4.fsdp_sft_trainer.py
位置:https://github.com/volcengine/verl/blob/main/verl/trainer/fsdp_sft_trainer.py
修改点:_compute_loss_and_backward 函数的 loss.backward()往前移动4格。
```

#### 拉起命令

```shell
# 具体根据verl-main目录修改
cd /public/home/fugx1/ds/verl-main/examples/sft/gsm8k

# 执行命令
bash run_deepseek_671b.sh 8 /public/home/fugx1/ds/verl-main/examples/sft/gsm8k
```

#### 脚本内容

```shell
# 需要根据实际路径修改
# data.train_files
# data.val_files
# model.partial_pretrain
# 环境变量根据实际情况修改

set -x

if [ "$#" -lt 2 ]; then
    echo "Usage: run_deepseek_6b7.sh <nproc_per_node> <save_path> [other_configs...]"
    exit 1
fi

nproc_per_node=$1
save_path=$2

source /opt/dtk/env.sh
export NCCL_P2P_LEVEL=PXB # SYS

# Runs Mixtral 8x7B model
export HIP_DIRECT_DISPATCH=0
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export GPU_MAX_HW_QUEUES=10
#export NVTE_FLASH_ATTN_TRITON=1
export NCCL_ALGO=Ring
export NCCL_SOCKET_IFNAME=enp33s0f3u1

export NCCL_NCHANNELS_PER_PEER=16
export NCCL_MIN_NCHANNELS=32 # 20
export NCCL_MAX_NCHANNELS=32 # 20
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1

export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
export NCCL_NET_GDR_LEVEL=7
export NCCL_NET_GDR_READ=1
export RCCL_SDMA_COPY_ENABLE=0
export NCCL_TOPO_FILE="/public/home/fugx1/datasets/rccl-test/topo-input.xml"
# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
export GLOG_minloglevel=3 # 打印error级别的nccl日志

export PATH=/opt/hpc/software/mpi/hpcx/2.12.0/gcc-8.3.1/bin/:$PATH
export LD_LIBRARY_PATH=/opt/hpc/software/mpi/hpcx/2.12.0/gcc-8.3.1/lib/:$LD_LIBRARY_PATH

# 导入hipblaslt库
export LD_LIBRARY_PATH=/public/home/fugx1/tests1/test03/whl/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH

# 更新rocblas
export LD_LIBRARY_PATH=/public/home/fugx1/tests1/test03/whl/rocblas-install-0224/lib:$LD_LIBRARY_PATH


RANK=$OMPI_COMM_WORLD_RANK
LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
WORLD_SIZE=$OMPI_COMM_WORLD_SIZE


# Shift the arguments so $@ refers to the rest
shift 2

torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \
     -m verl.trainer.fsdp_sft_trainer \
    data.train_files=/public/home/fugx1/ds/gsm8k/gsm8k-train.parquet \
    data.val_files=/public/home/fugx1/ds/gsm8k/gsm8k-test.parquet \
    data.prompt_key='question' \
    data.response_key='answer' \
    +data.prompt_dict_keys=['question'] \
    +data.response_dict_keys=['answer'] \
    data.micro_batch_size_per_gpu=1 \
    model.partial_pretrain=/public/opendas/DL_DATA/llm-models/DeepSeek-V3-bf16 \
    trainer.default_local_dir=$save_path \
    trainer.project_name=gsm8k-sft  \
    trainer.experiment_name=gsm8k-sft-deepseek-v3-671b \
    trainer.total_epochs=1 \
    trainer.logger=['console'] \
    trainer.default_hdfs_dir=null $@

```