K100AI_finetune.sh

set -eux
# 多节点环境变量
# Runs the "7B" parameter model
export HSA_FORCE_FINE_GRAIN_PCIE=1
export OMP_NUM_THREADS=1
export NCCL_P2P_LEVEL=5
source /opt/dtk/env.sh
# te调用gemm需要导入hipblaslt库
# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 

#export HIP_ALLOC_INITIALIZE=0
#export GPU_MAX_HW_QUEUES=20
export NCCL_ALGO=Ring
export NCCL_NCHANNELS_PER_PEER=8
export NCCL_MIN_NCHANNELS=20
export NCCL_MIN_P2P_NCHANNELS=8
export NCCL_IB_TIMEOUT=22
export CUDA_DEVICE_MAX_CONNECTIONS=1

export NCCL_IB_HCA=mlx5_1,mlx5_2
#export NCCL_SOCKET_IFNAME=ibs8
export NCCL_NET_GDR_LEVEL=SYS
export NCCL_NET_GDR_READ=0
#export NCCL_DEBUG=info

# 模型和数据集参数
MODEL="/data/model_weights/llama2_7b_nemo/llama2-7b.nemo"
TRAIN_DS="[/data/datasets/mlperf_llama/databricks-dolly-15k/training.jsonl]"
VALID_DS="[/data/datasets/mlperf_llama/databricks-dolly-15k/validation.jsonl]"
TEST_DS="[/data/datasets/mlperf_llama/databricks-dolly-15k/test.jsonl]"
VALID_NAMES="[databricks-dolly-15k]"

# 微调数据集占比
# TRAIN_DS="[/path/to/dataset_1.jsonl,/path/to/dataset_2.jsonl]"
# CONCAT_SAMPLING_PROBS="[0.3,0.7]" # "[1]" # 只有一个数据集设置为1
CONCAT_SAMPLING_PROBS="[1]"

# 运行训练脚本
torchrun --nproc_per_node 8 \
   /workspace/nemo_main/NeMo-2.0.0.rc0.beta/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py \
   trainer.precision=bf16 \
   trainer.devices=8 \
   trainer.num_nodes=1 \
   trainer.val_check_interval=15 \
   trainer.max_steps=300 \
   model.restore_from_path=${MODEL} \
   model.micro_batch_size=1 \
   model.global_batch_size=60 \
   model.tensor_model_parallel_size=2 \
   model.pipeline_model_parallel_size=2 \
   model.megatron_amp_O2=True \
   model.sequence_parallel=True \
   model.activations_checkpoint_granularity=selective \
   model.activations_checkpoint_method=uniform \
   model.optim.name=fused_adam \
   model.optim.lr=5e-6 \
   model.answer_only_loss=True \
   model.peft.peft_scheme=lora \
   model.data.train_ds.file_names=${TRAIN_DS} \
   model.data.validation_ds.file_names=${VALID_DS} \
   model.data.test_ds.file_names=${TEST_DS} \
   model.data.train_ds.concat_sampling_probabilities=${CONCAT_SAMPLING_PROBS} \
   model.data.train_ds.max_seq_length=4096 \
   model.data.validation_ds.max_seq_length=4096 \
   model.data.train_ds.num_workers=0 \
   model.data.validation_ds.num_workers=0 \
   model.data.test_ds.num_workers=0 \
   ++cluster_type=BCP