verl/examples/grpo_trainer/run_seed_oss_36b_fsdp.sh at main · aoshen02/verl · 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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
set -x

# ---- user-adjustable ----
MODEL_PATH=${MODEL_PATH:-ByteDance-Seed/Seed-OSS-36B-Base}
TRAIN_FILE=${TRAIN_FILE:-$HOME/data/gsm8k/train.parquet}
TEST_FILE=${TEST_FILE:-$HOME/data/gsm8k/test.parquet}
NNODES=${NNODES:-1}
NGPUS_PER_NODE=${NGPUS_PER_NODE:-8}

TRAIN_BATCH_SIZE=${TRAIN_BATCH_SIZE:-64}
PPO_MINI_BATCH_SIZE=${PPO_MINI_BATCH_SIZE:-8}
PPO_MICRO_BATCH_SIZE_PER_GPU=${PPO_MICRO_BATCH_SIZE_PER_GPU:-2}
LOG_PROB_MICRO_BATCH_SIZE_PER_GPU=${LOG_PROB_MICRO_BATCH_SIZE_PER_GPU:-2}
MAX_PROMPT_LENGTH=${MAX_PROMPT_LENGTH:-512}
MAX_RESPONSE_LENGTH=${MAX_RESPONSE_LENGTH:-1024}

ACTOR_LR=${ACTOR_LR:-1e-6}
KL_LOSS_COEF=${KL_LOSS_COEF:-0.001}
ENTROPY_COEFF=${ENTROPY_COEFF:-0}

ROLLOUT_TP=${ROLLOUT_TP:-2}
ROLLOUT_GPU_MEM_UTIL=${ROLLOUT_GPU_MEM_UTIL:-0.6}
ROLLOUT_N=${ROLLOUT_N:-2}

PROJECT_NAME=${PROJECT_NAME:-verl_grpo_seed_oss_36b}
EXPERIMENT_NAME=${EXPERIMENT_NAME:-seed_oss_36b}
SAVE_FREQ=${SAVE_FREQ:-20}
TEST_FREQ=${TEST_FREQ:-5}
TOTAL_EPOCHS=${TOTAL_EPOCHS:-15}
# ---- end user-adjustable ----
########################### parameter arrays ###########################

DATA=(
    algorithm.adv_estimator=grpo
    data.train_files=${TRAIN_FILE}
    data.val_files=${TEST_FILE}
    data.train_batch_size=${TRAIN_BATCH_SIZE}
    data.max_prompt_length=${MAX_PROMPT_LENGTH}
    data.max_response_length=${MAX_RESPONSE_LENGTH}
    data.filter_overlong_prompts=True
    data.truncation='error'
    algorithm.use_kl_in_reward=False
)

MODEL=(
    actor_rollout_ref.model.path=${MODEL_PATH}
    actor_rollout_ref.model.use_remove_padding=True
    actor_rollout_ref.model.enable_gradient_checkpointing=True
    actor_rollout_ref.model.use_fused_kernels=True
)

ACTOR=(
    actor_rollout_ref.actor.optim.lr=${ACTOR_LR}
    actor_rollout_ref.actor.ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE}
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${PPO_MICRO_BATCH_SIZE_PER_GPU}
    actor_rollout_ref.actor.use_kl_loss=True
    actor_rollout_ref.actor.kl_loss_coef=${KL_LOSS_COEF}
    actor_rollout_ref.actor.kl_loss_type=low_var_kl
    actor_rollout_ref.actor.entropy_coeff=${ENTROPY_COEFF}
    actor_rollout_ref.actor.use_dynamic_bsz=True
    actor_rollout_ref.actor.strategy=fsdp2
    actor_rollout_ref.actor.fsdp_config.param_offload=True
    actor_rollout_ref.actor.fsdp_config.param_offload=True
)

ROLLOUT=(
    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=True
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=${LOG_PROB_MICRO_BATCH_SIZE_PER_GPU}
    actor_rollout_ref.rollout.tensor_model_parallel_size=${ROLLOUT_TP}
    actor_rollout_ref.rollout.name=vllm
    actor_rollout_ref.rollout.gpu_memory_utilization=${ROLLOUT_GPU_MEM_UTIL}
    actor_rollout_ref.rollout.n=${ROLLOUT_N}
    actor_rollout_ref.rollout.free_cache_engine=True
)

REF=(
    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=True
    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=${LOG_PROB_MICRO_BATCH_SIZE_PER_GPU}
    actor_rollout_ref.ref.fsdp_config.param_offload=True
    actor_rollout_ref.ref.strategy=fsdp2
)

TRAINER=(
    trainer.critic_warmup=0
    trainer.logger='["console","wandb"]'
    trainer.project_name=${PROJECT_NAME}
    trainer.experiment_name=${EXPERIMENT_NAME}
    trainer.val_before_train=False
    trainer.n_gpus_per_node=${NGPUS_PER_NODE}
    trainer.nnodes=${NNODES}
    trainer.save_freq=${SAVE_FREQ}
    trainer.test_freq=${TEST_FREQ}
    trainer.total_epochs=${TOTAL_EPOCHS}
)

EXTRA=(
)

########################### launch ###########################
python3 -m verl.trainer.main_ppo \
    "${DATA[@]}" \
    "${MODEL[@]}" \
    "${ACTOR[@]}" \
    "${ROLLOUT[@]}" \
    "${REF[@]}" \
    "${TRAINER[@]}" \
    "${EXTRA[@]}" \
    "$@"