mirror of
https://github.com/hanzoai/open-instruct.git
synced 2026-10-30 18:26:32 +00:00
Costa HuangandGitHub
28be6c8402
* use the latest oe-eval-image * use the latest image * rename * do not use slow tokenizer
73 lines
2.1 KiB
YAML
73 lines
2.1 KiB
YAML
version: v2
|
|
description: open-instruct-finetune-multinode
|
|
budget: ai2/oe-adapt
|
|
tasks:
|
|
- name: open-instruct-finetune-multinode
|
|
replicas: 4
|
|
leaderSelection: true
|
|
hostNetworking: true
|
|
propagateFailure: true
|
|
propagatePreemption: true
|
|
synchronizedStartTimeout: 60m
|
|
image:
|
|
beaker: nathanl/open_instruct_auto
|
|
command: [
|
|
'/bin/sh', '-c'
|
|
]
|
|
arguments: ['
|
|
unset CUDA_LAUNCH_BLOCKING && PYTHONPATH="/stage:$PYTHONPATH" pip install git+https://github.com/vwxyzjn/transformers.git@olmo2-classification && 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
|
|
--use_deepspeed
|
|
--deepspeed_config_file configs/ds_configs/stage3_no_offloading_accelerate.conf
|
|
--deepspeed_multinode_launcher standard
|
|
open_instruct/dpo_tune_cache.py
|
|
--model_name_or_path /hf_llama_models
|
|
--use_flash_attn
|
|
--max_seq_length 4096
|
|
--preprocessing_num_workers 16
|
|
--per_device_train_batch_size 1
|
|
--gradient_accumulation_steps 4
|
|
--learning_rate 5e-7
|
|
--lr_scheduler_type linear
|
|
--warmup_ratio 0.1
|
|
--weight_decay 0.
|
|
--num_train_epochs 3
|
|
--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_API_KEY
|
|
# secret: WANDB_API_KEY
|
|
- name: WANDB_PROJECT
|
|
value: open-instruct
|
|
- name: WANDB_WATCH
|
|
value: false
|
|
- name: WANDB_LOG_MODEL
|
|
value: false
|
|
- name: WANDB_DISABLED
|
|
value: true
|
|
# - name: HF_TOKEN
|
|
# secret: HF_TOKEN
|
|
# datasets:
|
|
# - mountPath: /oe-adapt-default
|
|
# source:
|
|
# weka: oe-adapt-default
|
|
result:
|
|
path: /output
|
|
resources:
|
|
gpuCount: 8
|
|
context:
|
|
cluster: ai2/allennlp-cirrascale
|
|
priority: high
|
|
preemptible: false |