verl/recipe/flash_rl/gsm8k_qwen0_5b_int8.sh at flash-rl · yaof20/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
RUN_NAME=${1:-"gsm8k-PPO-Qwen2.5-0.5B-w8a8-on-policy"}
CAP=${2:-"-1"}
FP32_LM_HEAD=${3:-"0"}

project_name='GSM8K-PPO'
exp_name=${RUN_NAME}

export VERL_LOGGING_LEVEL=DEBUG
export VLLM_LOGGING_LEVEL=DEBUG
export VLLM_CONFIGURE_LOGGING=1
export FLASHRL_LOGGING_LEVEL=DEBUG
export FLASHRL_CONFIG=LiyuanLucasLiu/Qwen2.5-0.5B-Instruct-quantized.w8a8-RedHatAI/flashrl_config.yaml
export FLASHRL_LMHEAD_FP32=${FP32_LM_HEAD}

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=gae \
    data.train_files=$HOME/data/gsm8k/train.parquet \
    data.val_files=$HOME/data/gsm8k/test.parquet \
    data.train_batch_size=256 \
    data.max_prompt_length=1024 \
    data.max_response_length=512 \
    data.filter_overlong_prompts=True \
    data.truncation='error' \
    actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B-Instruct \
    actor_rollout_ref.actor.optim.lr=1e-6 \
    actor_rollout_ref.model.use_remove_padding=False \
    actor_rollout_ref.actor.ppo_mini_batch_size=64 \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.actor.fsdp_config.param_offload=False \
    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
    actor_rollout_ref.actor.use_kl_loss=False \
    actor_rollout_ref.actor.imp_ratio_cap=${CAP} \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \
    actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
    actor_rollout_ref.rollout.name=vllm \
    actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \
    actor_rollout_ref.rollout.disable_log_stats=False \
    critic.optim.lr=1e-5 \
    critic.model.use_remove_padding=False \
    critic.model.path=Qwen/Qwen2.5-0.5B-Instruct  \
    critic.model.enable_gradient_checkpointing=False \
    critic.ppo_micro_batch_size_per_gpu=4 \
    critic.model.fsdp_config.param_offload=False \
    critic.model.fsdp_config.optimizer_offload=False \
    algorithm.use_kl_in_reward=False \
    trainer.critic_warmup=0 \
    trainer.logger=['console','wandb'] \
    trainer.project_name="${project_name}" \
    trainer.experiment_name="${exp_name}" \
    trainer.n_gpus_per_node=4 \
    trainer.val_before_train=True \
    trainer.nnodes=1 \
    trainer.save_freq=20 \
    trainer.test_freq=10 \
    trainer.total_epochs=15 2>&1 | tee $RUN_NAME.log