Basic curriculum (#198)

* feat: Add optional curriculum support to dataset registration and creation
* docs: Add docstrings to create_curriculum() and register_dataset()
* feat: Add curriculum configuration classes for CurriculumExperiment
* feat: Add weight parameter to CurriculumAttributeConfig and use in DatasetSpec
* refactor: Simplify CurriculumAttributeConfig with "*" attribute level support
* test: Add unit tests for CurriculumExperiment class
* feat: Add from_yaml() method to CurriculumExperimentConfig with unit test

examples/veRL/basic_curriculum/config/ppo_trainer.yaml

@@ -0,0 +1,171 @@
+data:
+  tokenizer: null
+  train_files: ~/data/rlhf/gsm8k/train.parquet
+  val_files: ~/data/rlhf/gsm8k/test.parquet
+  prompt_key: prompt
+  max_prompt_length: 512
+  max_response_length: 512
+  train_batch_size: 1024
+  val_batch_size: 1312
+  return_raw_input_ids: False  # This should be set to true when the tokenizer between policy and rm differs
+  return_raw_chat: False
+
+actor_rollout_ref:
+  hybrid_engine: True
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    external_lib: null
+    override_config: { }
+    enable_gradient_checkpointing: True
+    use_remove_padding: False
+  actor:
+    strategy: fsdp  # This is for backward-compatibility
+    ppo_mini_batch_size: 256
+    ppo_micro_batch_size: null # will be deprecated, use ppo_micro_batch_size_per_gpu
+    ppo_micro_batch_size_per_gpu: null
+    use_dynamic_bsz: False
+    ppo_max_token_len_per_gpu: 16384 # n * ${data.max_prompt_length} + ${data.max_response_length}
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    entropy_coeff: 0.001
+    use_kl_loss: False # True for GRPO
+    kl_loss_coef: 0.001 # for grpo
+    kl_loss_type: low_var_kl # for grpo
+    ppo_epochs: 1
+    shuffle: False
+    ulysses_sequence_parallel_size: 1 # sp size
+    optim:
+      lr: 1e-6
+      lr_warmup_steps_ratio: 0.  # the total steps will be injected during runtime
+      min_lr_ratio: null   # only useful for warmup with cosine
+      warmup_style: constant  # select from constant/cosine
+      total_training_steps: -1  # must be override by program
+    fsdp_config:
+      wrap_policy:
+        # transformer_layer_cls_to_wrap: None
+        min_num_params: 0
+      param_offload: False
+      optimizer_offload: False
+      fsdp_size: -1
+  ref:
+    fsdp_config:
+      param_offload: False
+      wrap_policy:
+        # transformer_layer_cls_to_wrap: None
+        min_num_params: 0
+    log_prob_micro_batch_size: null # will be deprecated, use log_prob_micro_batch_size_per_gpu
+    log_prob_micro_batch_size_per_gpu: null
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    ulysses_sequence_parallel_size: ${actor_rollout_ref.actor.ulysses_sequence_parallel_size} # sp size
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1 # 0 for hf rollout, -1 for vllm rollout
+    top_p: 1
+    prompt_length: ${data.max_prompt_length}  # not use for opensource
+    response_length: ${data.max_response_length}
+    # for vllm rollout
+    dtype: bfloat16 # should align with FSDP
+    gpu_memory_utilization: 0.5
+    ignore_eos: False
+    enforce_eager: True
+    free_cache_engine: True
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 2
+    max_num_batched_tokens: 8192
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null # will be deprecated, use log_prob_micro_batch_size_per_gpu
+    log_prob_micro_batch_size_per_gpu: null
+    log_prob_use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+    log_prob_max_token_len_per_gpu: ${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}
+    disable_log_stats: True
+    enable_chunked_prefill: True # could get higher throughput
+    # for hf rollout
+    do_sample: True
+    # number of responses (i.e. num sample times)
+    n: 1 # > 1 for grpo
+
+critic:
+  strategy: fsdp
+  optim:
+    lr: 1e-5
+    lr_warmup_steps_ratio: 0.  # the total steps will be injected during runtime
+    min_lr_ratio: null   # only useful for warmup with cosine
+    warmup_style: constant  # select from constant/cosine
+    total_training_steps: -1  # must be override by program
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: ${actor_rollout_ref.model.path}
+    override_config: { }
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    enable_gradient_checkpointing: True
+    use_remove_padding: False
+    fsdp_config:
+      param_offload: False
+      optimizer_offload: False
+      wrap_policy:
+        # transformer_layer_cls_to_wrap: None
+        min_num_params: 0
+      fsdp_size: -1
+  ppo_mini_batch_size: ${actor_rollout_ref.actor.ppo_mini_batch_size}
+  ppo_micro_batch_size: null # will be deprecated, use ppo_micro_batch_size_per_gpu
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: ${critic.ppo_micro_batch_size}
+  forward_micro_batch_size_per_gpu: ${critic.ppo_micro_batch_size_per_gpu}
+  use_dynamic_bsz: ${actor_rollout_ref.actor.use_dynamic_bsz}
+  ppo_max_token_len_per_gpu: 32768 # (${actor_rollout_ref.actor.ppo_max_token_len_per_gpu}) * 2
+  forward_max_token_len_per_gpu: ${critic.ppo_max_token_len_per_gpu}
+  ulysses_sequence_parallel_size: 1 # sp size
+  ppo_epochs: ${actor_rollout_ref.actor.ppo_epochs}
+  shuffle: ${actor_rollout_ref.actor.shuffle}
+  grad_clip: 1.0
+  cliprange_value: 0.5
+
+reward_model:
+  enable: False
+  strategy: fsdp
+  model:
+    input_tokenizer: ${actor_rollout_ref.model.path}  # set this to null if the chat template is identical
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: ${actor_rollout_ref.model.external_lib}
+    use_remove_padding: False
+    fsdp_config:
+      min_num_params: 0
+      param_offload: False
+      fsdp_size: -1
+  micro_batch_size: null # will be deprecated, use micro_batch_size_per_gpu
+  micro_batch_size_per_gpu: null # set a number
+  max_length: null
+  ulysses_sequence_parallel_size: 1 # sp size
+  use_dynamic_bsz: ${critic.use_dynamic_bsz}
+  forward_max_token_len_per_gpu: ${critic.forward_max_token_len_per_gpu}
+
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  kl_penalty: kl  # how to estimate kl divergence
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+
+trainer:
+  total_epochs: 30
+  total_training_steps: null
+  project_name: verl_examples
+  experiment_name: gsm8k
+  logger: [ 'console', 'wandb' ]
+  val_generations_to_log_to_wandb: 0
+  nnodes: 1
+  n_gpus_per_node: 8
+  save_freq: -1
+  # auto: find the last ckpt to resume. If can't find, start from scratch
+  resume_mode: auto # or auto or resume_path if
+  resume_from_path: False
+  test_freq: -1
+  critic_warmup: 0
+  default_hdfs_dir: null
+  remove_previous_ckpt_in_save: False
+  del_local_ckpt_after_load: False
+  default_local_dir: checkpoints/${trainer.project_name}/${trainer.experiment_name}

examples/veRL/basic_curriculum/launch.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+
+export N_GPUS=4
+export BASE_MODEL=meta-llama/Llama-3.2-3B-Instruct  # meta-llama/Llama-3.2-1B-Instruct
+export ROLLOUT_TP_SIZE=2
+export EXPERIMENT_NAME=basic_curriculum
+export VLLM_ATTENTION_BACKEND=XFORMERS
+
+bash ./train_grpo.sh

examples/veRL/basic_curriculum/ppo_curriculum.py

@@ -0,0 +1,298 @@
+# This example is an adapted version of Bytedance's code:
+# https://github.com/volcengine/verl/blob/a65c9157bc0b85b64cd753de19f94e80a11bd871/verl/trainer/main_ppo.py
+from io import StringIO
+from typing import Optional
+
+import hydra
+import ray
+import torch
+import verl.utils.torch_functional as verl_F
+from omegaconf import OmegaConf, open_dict
+from torch.utils.data import DataLoader, Dataset
+from transformers import PreTrainedTokenizer
+from verl import DataProto
+from verl.trainer.ppo.ray_trainer import RayPPOTrainer
+from verl.utils.dataset.rl_dataset import collate_fn
+from verl.utils.model import compute_position_id_with_mask
+
+import reasoning_gym
+import reasoning_gym.utils
+from reasoning_gym.coaching.curriculum_config import CurriculumExperimentConfig
+from reasoning_gym.coaching.experiment import CurriculumExperiment
+from reasoning_gym.utils import extract_answer
+
+curriculum_config_yaml = """
+ curricula:
+   leg_counting:
+     attribute_levels:
+       num_animals: 2
+   products:
+     attribute_levels:
+       num_terms: 4
+       num_digits: 4
+   chain_sum:
+     attribute_levels:
+       num_terms: 4
+       num_digits: 4
+     weight: 1.0
+ """
+
+
+class ReasoningGymDataset(Dataset):
+    def __init__(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        experiment_name: str,
+        seed: int,
+        size: int,
+        developer_prompt: Optional[str] = None,
+        developer_role: str = "system",
+        max_prompt_length: int = 2048,
+        truncation: str = "error",  ##  ['left', 'right', 'error']
+        return_raw_chat: bool = False,
+    ):
+        self.tokenizer = tokenizer
+        curriculum_config = CurriculumExperimentConfig.from_yaml_stream(StringIO(curriculum_config_yaml))
+        self.experiment = CurriculumExperiment(experiment_name, curriculum_config, size=size, seed=seed)
+        self.developer_prompt = developer_prompt
+        self.developer_role = developer_role
+        self.max_prompt_length = max_prompt_length
+        self.truncation = truncation
+        self.return_raw_chat = return_raw_chat
+
+    def __len__(self) -> int:
+        return len(self.experiment.composite)
+
+    def __getitem__(self, index: int):
+        row_dict = self.experiment.get_dataset_entry(index).copy()
+        q = row_dict["question"]
+
+        chat = []
+        if self.developer_prompt is not None:
+            chat.append({"role": self.developer_role, "content": self.developer_prompt})
+        chat.append({"role": "user", "content": q})
+
+        prompt = self.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
+
+        input_ids, attention_mask = verl_F.tokenize_and_postprocess_data(
+            prompt=prompt,
+            tokenizer=self.tokenizer,
+            max_length=self.max_prompt_length,
+            pad_token_id=self.tokenizer.pad_token_id,
+            left_pad=True,
+            truncation=self.truncation,
+        )
+
+        position_ids = compute_position_id_with_mask(attention_mask)
+
+        row_dict["data_source"] = "reasoning_gym/" + self.dataset_name
+        row_dict["input_ids"] = input_ids[0]
+        row_dict["attention_mask"] = attention_mask[0]
+        row_dict["position_ids"] = position_ids[0]
+
+        # encode prompts without chat template
+        if self.return_raw_chat:
+            row_dict["raw_prompt"] = chat.tolist()
+
+        return row_dict
+
+
+class RayPPOTrainerCustom(RayPPOTrainer):
+    def __init__(
+        self,
+        config,
+        tokenizer,
+        role_worker_mapping: dict,
+        resource_pool_manager,
+        ray_worker_group_cls,
+        experiment_name: str = "basic_curriculum",
+        dataset_size: int = 10000,
+    ):
+        self.dataset_size = dataset_size
+
+        developer_prompt = reasoning_gym.utils.SYSTEM_PROMPTS["DeepSeekZero"]
+        self.train_dataset = ReasoningGymDataset(
+            tokenizer=tokenizer,
+            experiment_name=experiment_name,
+            seed=1,
+            size=self.dataset_size,
+            developer_prompt=developer_prompt,
+        )
+
+        self.val_dataset = ReasoningGymDataset(
+            tokenizer=tokenizer,
+            experiment_name=experiment_name,
+            seed=2,
+            size=self.dataset_size,
+            developer_prompt=developer_prompt,
+        )
+
+        train_reward_fn = lambda data: self._score_output(data, num_examine=0)
+        val_reward_fn = lambda data: self._score_output(data, num_examine=1)
+
+        super().__init__(
+            config,
+            tokenizer,
+            role_worker_mapping,
+            resource_pool_manager,
+            ray_worker_group_cls,
+            train_reward_fn,
+            val_reward_fn,
+        )
+
+    def _score_output(self, data: DataProto, num_examine: int = 0) -> torch.Tensor:
+        reward_tensor = torch.zeros_like(data.batch["responses"], dtype=torch.float32)
+
+        num_printed = 0
+        for i in range(len(data)):
+            data_item = data[i]  # DataProtoItem
+
+            prompt_ids = data_item.batch["prompts"]  # tokenized 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)
+
+            entry_id = data_item.non_tensor_batch["metadata"]["entry_id"]
+
+            score = self._compute_score(
+                solution_str=sequences_str,
+                entry_id=entry_id,
+            )
+            reward_tensor[i, valid_response_length - 1] = score
+
+            if num_printed < num_examine:
+                print(f"reward={score}, seq={sequences_str}")
+                num_printed += 1
+
+        return reward_tensor
+
+    def _compute_score(self, solution_str: str, entry_id: str) -> float:
+        found_answer = extract_answer(solution_str, tag_name="answer")
+        reward = self.train_dataset.experiment.score_answer_with_id(found_answer, entry_id=entry_id)
+        print(f"entry_id: {entry_id}; found answer={found_answer}; reward: {reward};")
+        return reward
+
+    def _create_dataloader(self):
+        self.train_dataloader = DataLoader(
+            dataset=self.train_dataset,
+            batch_size=self.config.data.train_batch_size,
+            shuffle=True,
+            drop_last=True,
+            collate_fn=collate_fn,
+        )
+
+        self.val_dataloader = DataLoader(
+            dataset=self.val_dataset,
+            batch_size=len(self.val_dataset),
+            shuffle=True,
+            drop_last=True,
+            collate_fn=collate_fn,
+        )
+
+        assert len(self.train_dataloader) >= 1
+        assert len(self.val_dataloader) >= 1
+
+        print(f"Size of train dataloader: {len(self.train_dataloader)}")
+        print(f"Size of val dataloader: {len(self.val_dataloader)}")
+
+        # inject total_training_steps to actor/critic optim_config. This is hacky.
+        total_training_steps = len(self.train_dataloader) * self.config.trainer.total_epochs
+
+        if self.config.trainer.total_training_steps is not None:
+            total_training_steps = self.config.trainer.total_training_steps
+
+        self.total_training_steps = total_training_steps
+        print(f"Total training steps: {self.total_training_steps}")
+
+        OmegaConf.set_struct(self.config, True)
+        with open_dict(self.config):
+            self.config.actor_rollout_ref.actor.optim.total_training_steps = total_training_steps
+            self.config.critic.optim.total_training_steps = total_training_steps
+
+
+@ray.remote
+def main_task(config):
+    # print initial config
+    from pprint import pprint
+
+    from verl.utils import hf_tokenizer
+    from verl.utils.fs import copy_local_path_from_hdfs
+
+    pprint(OmegaConf.to_container(config, resolve=True))  # resolve=True will eval symbol values
+    OmegaConf.resolve(config)
+
+    # download the checkpoint from hdfs
+    local_path = copy_local_path_from_hdfs(config.actor_rollout_ref.model.path)
+
+    # instantiate 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),
+        Role.RefPolicy: ray.remote(ActorRolloutRefWorker),
+    }
+
+    global_pool_id = "global_pool"
+    resource_pool_spec = {
+        global_pool_id: [config.trainer.n_gpus_per_node] * config.trainer.nnodes,
+    }
+    mapping = {
+        Role.ActorRollout: global_pool_id,
+        Role.Critic: global_pool_id,
+        Role.RefPolicy: global_pool_id,
+    }
+
+    resource_pool_manager = ResourcePoolManager(resource_pool_spec=resource_pool_spec, mapping=mapping)
+
+    trainer = RayPPOTrainerCustom(
+        config=config,
+        tokenizer=tokenizer,
+        role_worker_mapping=role_worker_mapping,
+        resource_pool_manager=resource_pool_manager,
+        ray_worker_group_cls=ray_worker_group_cls,
+    )
+    trainer.init_workers()
+    trainer.fit()
+
+
+@hydra.main(config_path="config", config_name="ppo_trainer", 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"}})
+
+    ray.get(main_task.remote(config))
+
+
+if __name__ == "__main__":
+    main()

examples/veRL/basic_curriculum/train_grpo.sh

@@ -0,0 +1,39 @@
+#!/bin/bash
+set -x
+
+python3 -u ppo_curriculum.py \
+    algorithm.adv_estimator=grpo \
+    data.train_files=$DATA_DIR/train.parquet \
+    data.val_files=$DATA_DIR/test.parquet \
+    data.train_batch_size=512 \
+    data.val_batch_size=512 \
+    data.max_prompt_length=512 \
+    data.max_response_length=1024 \
+    actor_rollout_ref.model.path=$BASE_MODEL \
+    actor_rollout_ref.actor.optim.lr=1e-6 \
+    actor_rollout_ref.model.use_remove_padding=True \
+    actor_rollout_ref.actor.ppo_mini_batch_size=256 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=80 \
+    actor_rollout_ref.actor.use_kl_loss=True \
+    actor_rollout_ref.actor.kl_loss_coef=0.001 \
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=160 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP_SIZE \
+    actor_rollout_ref.rollout.name=vllm \
+    actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \
+    actor_rollout_ref.rollout.n=8 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=160 \
+    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    algorithm.kl_ctrl.kl_coef=0.001 \
+    trainer.critic_warmup=0 \
+    trainer.logger=['wandb'] \
+    trainer.project_name='verl_chain_sum_grpo' \
+    trainer.experiment_name=$EXPERIMENT_NAME \
+    trainer.n_gpus_per_node=$N_GPUS \
+    trainer.nnodes=1 \
+    trainer.save_freq=100 \
+    trainer.test_freq=100 \
+    trainer.total_epochs=15 $@ 2>&1 | tee verl_output.log