mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Yizhong WangandGitHub
e96a951642
* Add parameters in multinode training to propagate failure/preemptation and set timeout.
78 lines
2.3 KiB
YAML
78 lines
2.3 KiB
YAML
version: v2
|
|
description: open-instruct-finetune-multinode-test
|
|
tasks:
|
|
- name: open-instruct-finetune-multinode-test
|
|
replicas: 4
|
|
leaderSelection: true
|
|
hostNetworking: true
|
|
propagateFailure: true
|
|
propagatePreemption: true
|
|
synchronizedStartTimeout: 15m
|
|
image:
|
|
beaker: Yizhongw03/open-instruct-multi-node
|
|
command: [
|
|
'/bin/sh', '-c'
|
|
]
|
|
arguments: ['
|
|
unset CUDA_LAUNCH_BLOCKING && accelerate launch
|
|
--mixed_precision bf16
|
|
--num_machines 4
|
|
--num_processes 32
|
|
--machine_rank $BEAKER_REPLICA_RANK
|
|
--main_process_ip $BEAKER_LEADER_REPLICA_HOSTNAME
|
|
--main_process_port 29400
|
|
open_instruct/finetune.py
|
|
--model_name_or_path /net/nfs.cirrascale/allennlp/yizhongw/hf_llama2_models/70B
|
|
--tokenizer_name /net/nfs.cirrascale/allennlp/yizhongw/hf_llama2_models/70B
|
|
--use_slow_tokenizer
|
|
--train_file /net/nfs.cirrascale/allennlp/yizhongw/open-instruct-public/data/processed/tulu/tulu_v1_mix.jsonl
|
|
--use_flash_attn
|
|
--use_lora
|
|
--use_qlora
|
|
--lora_rank 64
|
|
--lora_alpha 64
|
|
--lora_dropout 0.1
|
|
--gradient_checkpointing
|
|
--max_seq_length 2048
|
|
--preprocessing_num_workers 64
|
|
--per_device_train_batch_size 1
|
|
--gradient_accumulation_steps 4
|
|
--learning_rate 2e-5
|
|
--lr_scheduler_type linear
|
|
--warmup_ratio 0.03
|
|
--weight_decay 0.
|
|
--num_train_epochs 5
|
|
--output_dir /output/
|
|
--with_tracking
|
|
--report_to tensorboard
|
|
--logging_steps 1
|
|
']
|
|
envVars:
|
|
- name: CUDA_DEVICE_ORDER
|
|
value: PCI_BUS_ID
|
|
- name: TRANSFORMERS_CACHE
|
|
value: ./cache/
|
|
- name: WANDB_PROJECT
|
|
value: open-instruct
|
|
- name: WANDB_WATCH
|
|
value: false
|
|
- name: WANDB_LOG_MODEL
|
|
value: false
|
|
- name: WANDB_DISABLED
|
|
value: true
|
|
- name: NCCL_NET
|
|
value: IB
|
|
- name: NCCL_DEBUG
|
|
value: INFO
|
|
datasets:
|
|
- mountPath: /net/nfs.cirrascale
|
|
source:
|
|
hostPath: /net/nfs.cirrascale
|
|
result:
|
|
path: /output
|
|
resources:
|
|
gpuCount: 8
|
|
context:
|
|
priority: high
|
|
constraints:
|
|
cluster: [ai2/general-cirrascale-a100-80g-ib] |