File size: 7,984 Bytes
66407c5 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 |
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Note that we don't combine the main with ray_trainer as ray_trainer is used by other main.
"""
import hydra
import ray
import torch
from split_monkey_patch import fit
from verl import DataProto
from verl.trainer.ppo.ray_trainer import RayPPOTrainer
from verl.utils.reward_score import gsm8k, math
def _select_rm_score_fn(data_source):
if data_source == "openai/gsm8k":
return gsm8k.compute_score
elif data_source == "lighteval/MATH":
return math.compute_score
else:
raise NotImplementedError
class RewardManager:
def __init__(self, tokenizer, num_examine) -> None:
self.tokenizer = tokenizer
self.num_examine = num_examine # the number of batches of decoded responses to print to the console
def __call__(self, data: DataProto, return_dict: bool = False):
"""We will expand this function gradually based on the available datasets"""
# If there is rm score, we directly return rm score. Otherwise, we compute via rm_score_fn
if "rm_scores" in data.batch.keys():
return data.batch["rm_scores"]
reward_tensor = torch.zeros_like(data.batch["responses"], dtype=torch.float32)
already_print_data_sources = {}
for i in range(len(data)):
data_item = data[i] # DataProtoItem
prompt_ids = data_item.batch["prompts"]
prompt_length = prompt_ids.shape[-1]
valid_prompt_length = data_item.batch["attention_mask"][:prompt_length].sum()
valid_prompt_ids = prompt_ids[-valid_prompt_length:]
response_ids = data_item.batch["responses"]
valid_response_length = data_item.batch["attention_mask"][prompt_length:].sum()
valid_response_ids = response_ids[:valid_response_length]
# decode
sequences = torch.cat((valid_prompt_ids, valid_response_ids))
sequences_str = self.tokenizer.decode(sequences)
ground_truth = data_item.non_tensor_batch["reward_model"]["ground_truth"]
# select rm_score
data_source = data_item.non_tensor_batch["data_source"]
compute_score_fn = _select_rm_score_fn(data_source)
score = compute_score_fn(solution_str=sequences_str, ground_truth=ground_truth)
reward_tensor[i, valid_response_length - 1] = score
if data_source not in already_print_data_sources:
already_print_data_sources[data_source] = 0
if already_print_data_sources[data_source] < self.num_examine:
already_print_data_sources[data_source] += 1
print(sequences_str)
if return_dict:
return {"reward_tensor": reward_tensor}
else:
return reward_tensor
@hydra.main(config_path="config", config_name="ppo_trainer_split", version_base=None)
def main(config):
if not ray.is_initialized():
# this is for local ray cluster
ray.init(
runtime_env={"env_vars": {"TOKENIZERS_PARALLELISM": "true", "NCCL_DEBUG": "WARN"}},
num_cpus=config.ray_init.num_cpus,
)
ray.get(main_task.remote(config))
@ray.remote
def main_task(config):
# print initial config
from pprint import pprint
from omegaconf import OmegaConf
from verl.utils.fs import copy_to_local
pprint(OmegaConf.to_container(config, resolve=True)) # resolve=True will eval symbol values
OmegaConf.resolve(config)
# download the checkpoint from hdfs
local_path = copy_to_local(config.actor_rollout_ref.model.path)
# instantiate tokenizer
from verl.utils import hf_tokenizer
tokenizer = hf_tokenizer(local_path)
# define worker classes
if config.actor_rollout_ref.actor.strategy == "fsdp":
assert config.actor_rollout_ref.actor.strategy == config.critic.strategy
from verl.single_controller.ray import RayWorkerGroup
from verl.workers.fsdp_workers import ActorRolloutRefWorker, CriticWorker
ray_worker_group_cls = RayWorkerGroup
elif config.actor_rollout_ref.actor.strategy == "megatron":
assert config.actor_rollout_ref.actor.strategy == config.critic.strategy
from verl.single_controller.ray.megatron import NVMegatronRayWorkerGroup
from verl.workers.megatron_workers import ActorRolloutRefWorker, CriticWorker
ray_worker_group_cls = NVMegatronRayWorkerGroup
else:
raise NotImplementedError
from verl.trainer.ppo.ray_trainer import ResourcePoolManager, Role
role_worker_mapping = {
Role.ActorRollout: ray.remote(ActorRolloutRefWorker),
Role.Critic: ray.remote(CriticWorker),
}
# NOTE: initialze two resource pool
actor_rollout_ref_pool_id = "actor_rollout_ref_pool"
critic_pool_id = "critic_pool"
if config.trainer.nnodes // 2 == 0 and config.trainer.n_gpus_per_node // 2 > 0:
resource_pool_spec = {
actor_rollout_ref_pool_id: [config.trainer.n_gpus_per_node // 2] * config.trainer.nnodes,
critic_pool_id: [config.trainer.n_gpus_per_node // 2] * config.trainer.nnodes,
}
else:
resource_pool_spec = {
actor_rollout_ref_pool_id: [config.trainer.n_gpus_per_node] * (config.trainer.nnodes // 2),
critic_pool_id: [config.trainer.n_gpus_per_node] * (config.trainer.nnodes // 2),
}
print(f"resource_pool_spec: {resource_pool_spec}")
mapping = {
Role.ActorRollout: actor_rollout_ref_pool_id,
Role.Critic: critic_pool_id,
}
# use reference model
if config.algorithm.use_kl_in_reward or config.actor_rollout_ref.actor.use_kl_loss:
role_worker_mapping[Role.RefPolicy] = ray.remote(ActorRolloutRefWorker)
mapping[Role.RefPolicy] = actor_rollout_ref_pool_id
# we should adopt a multi-source reward function here
# - for rule-based rm, we directly call a reward score
# - for model-based rm, we call a model
# - for code related prompt, we send to a sandbox if there are test cases
# - finally, we combine all the rewards together
# - The reward type depends on the tag of the data
if config.reward_model.enable:
if config.reward_model.strategy == "fsdp":
from verl.workers.fsdp_workers import RewardModelWorker
elif config.reward_model.strategy == "megatron":
from verl.workers.megatron_workers import RewardModelWorker
else:
raise NotImplementedError
role_worker_mapping[Role.RewardModel] = ray.remote(RewardModelWorker)
mapping[Role.RewardModel] = critic_pool_id
reward_fn = RewardManager(tokenizer=tokenizer, num_examine=0)
# Note that we always use function-based RM for validation
val_reward_fn = RewardManager(tokenizer=tokenizer, num_examine=1)
resource_pool_manager = ResourcePoolManager(resource_pool_spec=resource_pool_spec, mapping=mapping)
RayPPOTrainer.fit = fit
trainer = RayPPOTrainer(
config=config,
tokenizer=tokenizer,
role_worker_mapping=role_worker_mapping,
resource_pool_manager=resource_pool_manager,
ray_worker_group_cls=ray_worker_group_cls,
reward_fn=reward_fn,
val_reward_fn=val_reward_fn,
device_name=config.trainer.device,
)
trainer.init_workers()
trainer.fit()
if __name__ == "__main__":
main()
|