NuRL/stage1_llama.sh at main · SalesforceAIResearch/NuRL · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
set -x
export WANDB_PROJECT=NuRL
export WANDB_EXP=Llama-Stage1-GRPO

export VLLM_ATTENTION_BACKEND=FLASH_ATTN
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:False"
export VLLM_USE_V1=1
export VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
export VLLM_ENGINE_ITERATION_TIMEOUT_S=100000000000
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

TRAIN_DATA=NuRL/train_data/stage1_grpo/train.parquet
VAL_DATA=NuRL/train_data/stage1_grpo/test.parquet
MODEL_PATH=meta-llama/Llama-3.2-3B-Instruct
EXPERIMENT_NAME=Llama-Stage1-GRPO
N_NODE=1

python -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    data.train_files=${TRAIN_DATA} \
    data.val_files=${VAL_DATA} \
    data.train_batch_size=1024 \
    data.max_prompt_length=1800 \
    data.max_response_length=9000 \
    data.truncation=right \
    actor_rollout_ref.model.path=$MODEL_PATH  \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=True \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=1 \
    actor_rollout_ref.actor.ppo_mini_batch_size=1024 \
    actor_rollout_ref.actor.use_dynamic_bsz=True \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=40000 \
    actor_rollout_ref.rollout.max_num_batched_tokens=40000 \
    actor_rollout_ref.actor.use_kl_loss=False \
    actor_rollout_ref.actor.kl_loss_coef=0 \
    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
    actor_rollout_ref.actor.entropy_coeff=0 \
    actor_rollout_ref.actor.clip_ratio_low=0.2 \
    actor_rollout_ref.actor.clip_ratio_high=0.28 \
    actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
    actor_rollout_ref.model.enable_gradient_checkpointing=True \
    actor_rollout_ref.actor.fsdp_config.param_offload=True \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.temperature=1.0 \
    actor_rollout_ref.rollout.top_p=1.0 \
    actor_rollout_ref.rollout.top_k=-1 \
    actor_rollout_ref.rollout.enable_chunked_prefill=True \
    actor_rollout_ref.rollout.n=16 \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \
    actor_rollout_ref.rollout.val_kwargs.do_sample=True \
    actor_rollout_ref.rollout.val_kwargs.temperature=0.7 \
    actor_rollout_ref.rollout.val_kwargs.top_p=1.0 \
    actor_rollout_ref.rollout.val_kwargs.n=8 \
    actor_rollout_ref.ref.fsdp_config.param_offload=True \
    algorithm.kl_ctrl.kl_coef=0.001 \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name='NuRL' \
    trainer.experiment_name=$EXPERIMENT_NAME \
    trainer.val_before_train=True \
    trainer.n_gpus_per_node=8 \
    trainer.nnodes=$N_NODE \
    trainer.save_freq=9 \
    trainer.test_freq=5 \
    trainer.balance_batch=False \
    trainer.default_hdfs_dir=null \
    trainer.total_epochs=100 "${@:1}"