I use
t1-3b as a base model which it is trained by using
t1-101k.
I use the grpo algorithm in
deepscaler.
1CUDA_VISIBLE_DEVICES=2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
2 algorithm.adv_estimator=grpo \
3 data.train_files=/root/deepscaler/data/train.parquet \
4 data.val_files=/root/deepscaler/data/aime.parquet \
5 data.train_batch_size=96 \
6 data.val_batch_size=288 \
7 data.max_prompt_length=1024 \
8 data.max_response_length=8192 \
9 actor_rollout_ref.model.path=/workspace/R1/model/t1-3B \
10 actor_rollout_ref.actor.optim.lr=1e-6 \
11 actor_rollout_ref.model.use_remove_padding=True \
12 actor_rollout_ref.actor.ppo_mini_batch_size=48 \
13 actor_rollout_ref.actor.ppo_micro_batch_size=48 \
14 actor_rollout_ref.actor.use_dynamic_bsz=True \
15 actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
16 actor_rollout_ref.actor.use_kl_loss=True \
17 actor_rollout_ref.actor.kl_loss_coef=0.001 \
18 actor_rollout_ref.actor.kl_loss_type=low_var_kl \
19 actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
20 actor_rollout_ref.model.enable_gradient_checkpointing=True \
21 actor_rollout_ref.actor.fsdp_config.param_offload=False \
22 actor_rollout_ref.actor.fsdp_config.grad_offload=False \
23 actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
24 actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
25 actor_rollout_ref.rollout.name=vllm \
26 actor_rollout_ref.rollout.temperature=0.6 \
27 actor_rollout_ref.rollout.val_temperature=0.6 \
28 actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
29 actor_rollout_ref.rollout.n=12 \
30 actor_rollout_ref.rollout.n_val=6 \
31 actor_rollout_ref.ref.fsdp_config.param_offload=True \
32 algorithm.kl_ctrl.kl_coef=0.001 \
33 trainer.critic_warmup=0 \
34 trainer.logger=['console','wandb'] \
35 trainer.project_name='t1-3b' \
36 trainer.experiment_name='t1-3b-grpo-8k' \
37 +trainer.val_before_train=True \
38 trainer.n_gpus_per_node=6 \
39 trainer.nnodes=1 \
40 trainer.save_freq=50 \
41 trainer.test_freq=50 \
42 trainer.default_hdfs_dir=null \
43 trainer.total_epochs=5
1CUDA_VISIBLE_DEVICES=2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
2 algorithm.adv_estimator=grpo \
3 data.train_files=/root/deepscaler/data/train.parquet \
4 data.val_files=/root/deepscaler/data/aime.parquet \
5 data.train_batch_size=48 \
6 data.val_batch_size=144 \
7 data.max_prompt_length=1024 \
8 data.max_response_length=16384 \
9 actor_rollout_ref.model.path=/workspace/R1/deepscaler/checkpoints/t1-3b/t1-3b-grpo-8k/actor/global_step_450 \
10 actor_rollout_ref.actor.optim.lr=1e-6 \
11 actor_rollout_ref.model.use_remove_padding=True \
12 actor_rollout_ref.actor.ppo_mini_batch_size=48 \
13 actor_rollout_ref.actor.ppo_micro_batch_size=48 \
14 actor_rollout_ref.actor.use_dynamic_bsz=True \
15 actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
16 actor_rollout_ref.actor.use_kl_loss=True \
17 actor_rollout_ref.actor.kl_loss_coef=0.001 \
18 actor_rollout_ref.actor.kl_loss_type=low_var_kl \
19 actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
20 actor_rollout_ref.model.enable_gradient_checkpointing=True \
21 actor_rollout_ref.actor.fsdp_config.param_offload=False \
22 actor_rollout_ref.actor.fsdp_config.grad_offload=False \
23 actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
24 actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
25 actor_rollout_ref.rollout.name=vllm \
26 actor_rollout_ref.rollout.temperature=0.6 \
27 actor_rollout_ref.rollout.val_temperature=0.6 \
28 actor_rollout_ref.rollout.gpu_memory_utilization=0.85 \
29 actor_rollout_ref.rollout.n=12 \
30 actor_rollout_ref.rollout.n_val=6 \
31 actor_rollout_ref.ref.fsdp_config.param_offload=True \
32 algorithm.kl_ctrl.kl_coef=0.001 \
33 trainer.critic_warmup=0 \
34 trainer.logger=['console','wandb'] \
35 trainer.project_name='t1-3b' \
36 trainer.experiment_name='t1-3b-grpo-16k' \
37 +trainer.val_before_train=True \
38 trainer.n_gpus_per_node=6 \
39 trainer.nnodes=1 \
40 trainer.save_freq=20 \
41 trainer.test_freq=20 \
42 trainer.default_hdfs_dir=null \
43 trainer.total_epochs=5
44
1CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python3 -m verl.trainer.main_ppo \
2 algorithm.adv_estimator=grpo \
3 data.train_files=/root/deepscaler/data/train.parquet \
4 data.val_files=/root/deepscaler/data/aime.parquet \
5 data.train_batch_size=16 \
6 data.val_batch_size=16 \
7 data.max_prompt_length=1024 \
8 data.max_response_length=31744 \
9 actor_rollout_ref.model.path=/workspace/R1/deepscaler/checkpoints/t1-3b/t1-3b-grpo-16k/actor/global_step_460 \
10 actor_rollout_ref.actor.optim.lr=1e-6 \
11 actor_rollout_ref.model.use_remove_padding=True \
12 actor_rollout_ref.actor.ppo_mini_batch_size=16 \
13 actor_rollout_ref.actor.ppo_micro_batch_size=16 \
14 actor_rollout_ref.actor.use_dynamic_bsz=True \
15 actor_rollout_ref.actor.ppo_max_token_len_per_gpu=32768 \
16 actor_rollout_ref.actor.use_kl_loss=True \
17 actor_rollout_ref.actor.kl_loss_coef=0.001 \
18 actor_rollout_ref.actor.kl_loss_type=low_var_kl \
19 actor_rollout_ref.actor.ulysses_sequence_parallel_size=1 \
20 actor_rollout_ref.model.enable_gradient_checkpointing=True \
21 actor_rollout_ref.actor.fsdp_config.param_offload=False \
22 actor_rollout_ref.actor.fsdp_config.grad_offload=False \
23 actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
24 actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
25 actor_rollout_ref.rollout.name=vllm \
26 actor_rollout_ref.rollout.temperature=0.6 \
27 actor_rollout_ref.rollout.val_temperature=0.6 \
28 actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
29 actor_rollout_ref.rollout.n=8 \
30 actor_rollout_ref.rollout.n_val=8 \
31 actor_rollout_ref.ref.fsdp_config.param_offload=True \
32 algorithm.kl_ctrl.kl_coef=0.001 \
33 trainer.critic_warmup=0 \
34 trainer.logger=['console','wandb'] \
35 trainer.project_name='t1-3b' \
36 trainer.experiment_name='t1-3b-grpo-32k' \
37 +trainer.val_before_train=True \
38 trainer.n_gpus_per_node=8 \
39 trainer.nnodes=1 \
40 trainer.save_freq=20 \
41 trainer.test_freq=20 \
42 trainer.default_hdfs_dir=null \
43 trainer.total_epochs=5
I will release a tiny vlm( t1-vl-grpo ).
