Update 0506

README.md

@@ -101,6 +101,13 @@ pip install "tensordict<0.6"
 pip install -e .
 ```
 
+### 数据集
+可根据下面的样例数据去构造自己的数据集
+
+- Text dataset: https://huggingface.co/datasets/hiyouga/math12k
+- Image-text dataset: https://huggingface.co/datasets/hiyouga/geometry3k
+- Multi-image-text dataset: https://huggingface.co/datasets/hiyouga/journeybench-multi-image-vqa
+
 ### GRPO 训练
 
 ```bash
@@ -114,7 +121,8 @@ python3 scripts/model_merger.py --local_dir path_to_your_actor_checkpoint
 ```
 
 > [!NOTE]
-
+> 如果无法连接到Hugging Face，请先安装`pip install -U huggingface_hub hf_transfer`，再在启动前增加 `export HF_ENDPOINT=https://hf-mirror.com`命令
+>
 > 如果您想使用SwanLab日志记录器，请考虑使用 `bash examples/qwen2_5_vl_7b_geo3k_swanlab.sh`.
 
 ## 自定义数据集
@@ -137,3 +145,17 @@ python3 scripts/model_merger.py --local_dir path_to_your_actor_checkpoint
 这些功能目前暂时禁用，我们计划在未来的更新中逐一修复。
 
 - 视觉语言模型目前不兼容 padding-free 训练和 DeepSpeed Ulysses并行方法。
+
+### 常见问题及解决办法
+
+> ValueError: Image features and image tokens do not match: tokens: 8192, features 9800
+
+增大`data.max_prompt_length`的数值或者减小`data.max_pixels` 的数值.
+
+> RuntimeError: CUDA Error: out of memory at /workspace/csrc/cumem_allocator.cpp:62
+
+减小`worker.rollout.gpu_memory_utilization`的数值并且确认开启 `worker.actor.offload.offload_params`.
+
+> RuntimeError: 0 active drivers ([]). There should only be one.
+
+在当前python环境下卸载 `deepspeed`.

examples/baselines/qwen2_5_vl_3b_clevr.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \
@@ -9,6 +13,7 @@ python3 -m verl.trainer.main \
     data.format_prompt=./examples/format_prompt/r1v_format.jinja \
     worker.actor.model.model_path=${MODEL_PATH} \
     worker.rollout.tensor_parallel_size=1 \
+    worker.reward.reward_type=sequential \
     worker.reward.reward_function=./examples/reward_function/r1v.py:compute_score \
     trainer.experiment_name=qwen2_5_vl_3b_clevr \
     trainer.n_gpus_per_node=2

examples/baselines/qwen2_5_vl_3b_geoqa8k.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \
@@ -9,6 +13,7 @@ python3 -m verl.trainer.main \
     data.format_prompt=./examples/format_prompt/r1v_format.jinja \
     worker.actor.model.model_path=${MODEL_PATH} \
     worker.rollout.tensor_parallel_size=1 \
+    worker.reward.reward_type=sequential \
     worker.reward.reward_function=./examples/reward_function/r1v.py:compute_score \
     trainer.experiment_name=qwen2_5_vl_3b_geoqa8k \
     trainer.n_gpus_per_node=8

examples/config.yaml

@@ -9,6 +9,7 @@ data:
   rollout_batch_size: 512
   val_batch_size: 1024
   format_prompt: ./examples/format_prompt/math_format.jinja
+  override_chat_template: null
   shuffle: true
   seed: 1
   max_pixels: 4194304
@@ -70,16 +71,17 @@ worker:
       offload_params: false
 
   reward:
-    reward_type: function
+    reward_type: batch
     reward_function: ./examples/reward_function/math.py:compute_score
 
 trainer:
-  total_episodes: 15
-  logger: ["console", "wandb"]
+  total_epochs: 15
+  max_steps: null
   project_name: easy_r1
   experiment_name: qwen2_5_7b_math_grpo
-  n_gpus_per_node: 8
+  logger: ["console", "wandb"]
   nnodes: 1
+  n_gpus_per_node: 8
   val_freq: 5  # -1 to disable
   val_before_train: true
   val_only: false

examples/qwen2_5_7b_math_grpo.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-7B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \
     config=examples/config.yaml \
-    data.train_files=hiyouga/math12k@train \
-    data.val_files=hiyouga/math12k@test \
-    worker.actor.model.model_path=${MODEL_PATH} \
-    trainer.experiment_name=qwen2_5_7b_math_grpo \
-    trainer.n_gpus_per_node=8
+    worker.actor.model.model_path=${MODEL_PATH}

examples/qwen2_5_vl_32b_geo3k_grpo.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-32B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen2_5_vl_3b_geo3k_grpo.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-3B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen2_5_vl_7b_geo3k_grpo.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen2_5_vl_7b_geo3k_reinforce.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen2_5_vl_7b_geo3k_swanlab.sh

+#!/bin/bash
+
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen2_5_vl_7b_multi_image.sh

+#!/bin/bash
 # REMINDER: this script uses test data split and should ONLY be used for debugging. DO NOT use for training.
 
 set -x
 
+export PYTHONUNBUFFERED=1
+
 MODEL_PATH=Qwen/Qwen2.5-VL-7B-Instruct  # replace it with your local file path
 
 python3 -m verl.trainer.main \

examples/qwen3_4b_math_grpo.sh

+#!/bin/bash
+
+set -x
+
+export PYTHONUNBUFFERED=1
+
+MODEL_PATH=Qwen/Qwen3-4B  # replace it with your local file path
+
+python3 -m verl.trainer.main \
+    config=examples/config.yaml \
+    data.max_response_length=4096 \
+    worker.actor.model.model_path=${MODEL_PATH} \
+    trainer.experiment_name=qwen3_4b_math_grpo

examples/reward_function/math.py

@@ -13,28 +13,34 @@
 # limitations under the License.
 
 import re
-from typing import Dict
+from typing import Dict, List
 
 from mathruler.grader import extract_boxed_content, grade_answer
 
 
-def format_reward(predict_str: str) -> float:
+def format_reward(predict: str) -> float:
     pattern = re.compile(r"<think>.*</think>.*\\boxed\{.*\}.*", re.DOTALL)
-    format_match = re.fullmatch(pattern, predict_str)
+    format_match = re.fullmatch(pattern, predict)
     return 1.0 if format_match else 0.0
 
 
-def accuracy_reward(predict_str: str, ground_truth: str) -> float:
-    answer = extract_boxed_content(predict_str)
+def accuracy_reward(predict: str, ground_truth: str) -> float:
+    answer = extract_boxed_content(predict)
     return 1.0 if grade_answer(answer, ground_truth) else 0.0
 
 
-def compute_score(predict_str: str, ground_truth: str, format_weight: float = 0.1) -> Dict[str, float]:
-    predict_str = re.sub(r"\s*(<|>|/)\s*", r"\1", predict_str)  # handle qwen2.5vl-32b format
-    format_score = format_reward(predict_str)
-    accuracy_score = accuracy_reward(predict_str, ground_truth)
-    return {
+def compute_score(predicts: List[str], ground_truths: List[str], format_weight: float = 0.1) -> List[Dict[str, float]]:
+    scores = []
+    for predict, ground_truth in zip(predicts, ground_truths):
+        predict = re.sub(r"\s*(<|>|/)\s*", r"\1", predict)  # handle qwen2.5vl-32b format
+        format_score = format_reward(predict)
+        accuracy_score = accuracy_reward(predict, ground_truth)
+        scores.append(
+            {
                 "overall": (1 - format_weight) * accuracy_score + format_weight * format_score,
                 "format": format_score,
                 "accuracy": accuracy_score,
             }
+        )
+
+    return scores

examples/reward_function/r1v.py

@@ -18,16 +18,16 @@ from typing import Dict
 from mathruler.grader import grade_answer
 
 
-def format_reward(predict_str: str) -> float:
+def format_reward(predict: str) -> float:
     pattern = re.compile(r"<think>.*?</think>\s*<answer>.*?</answer>", re.DOTALL)
-    format_match = re.fullmatch(pattern, predict_str)
+    format_match = re.fullmatch(pattern, predict)
     return 1.0 if format_match else 0.0
 
 
-def accuracy_reward(predict_str: str, ground_truth: str) -> float:
+def accuracy_reward(predict: str, ground_truth: str) -> float:
     try:
-        content_match = re.search(r"<answer>(.*?)</answer>", predict_str)
-        given_answer = content_match.group(1).strip() if content_match else predict_str.strip()
+        content_match = re.search(r"<answer>(.*?)</answer>", predict)
+        given_answer = content_match.group(1).strip() if content_match else predict.strip()
         if grade_answer(given_answer, ground_truth.strip()):
             return 1.0
 
@@ -37,9 +37,9 @@ def accuracy_reward(predict_str: str, ground_truth: str) -> float:
     return 0.0
 
 
-def compute_score(predict_str: str, ground_truth: str, format_weight: float = 0.5) -> Dict[str, float]:
-    format_score = format_reward(predict_str)
-    accuracy_score = accuracy_reward(predict_str, ground_truth)
+def compute_score(predict: str, ground_truth: str, format_weight: float = 0.5) -> Dict[str, float]:
+    format_score = format_reward(predict)
+    accuracy_score = accuracy_reward(predict, ground_truth)
     return {
         "overall": (1 - format_weight) * accuracy_score + format_weight * format_score,
         "format": format_score,

examples/runtime_env.yaml

@@ -3,6 +3,7 @@ excludes: ["/.git/"]
 env_vars:
   TOKENIZERS_PARALLELISM: "true"
   NCCL_DEBUG: "WARN"
-  VLLM_LOGGING_LEVEL: "INFO"
+  VLLM_LOGGING_LEVEL: "WARN"
   TORCH_NCCL_AVOID_RECORD_STREAMS: "1"
   PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:False"
+  PYTHONUNBUFFERED: "1"

verl/models/monkey_patch.py

@@ -20,7 +20,7 @@ from .transformers.qwen2_vl import qwen2_vl_attn_forward
 
 
 def apply_ulysses_patch(model_type: str) -> None:
-    if model_type in ("llama", "gemma", "gemma2", "mistral", "qwen2"):
+    if model_type in ("llama", "gemma", "gemma2", "mistral", "qwen2", "qwen3", "qwen3_moe"):
         ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward
     elif model_type in ("qwen2_vl", "qwen2_5_vl"):
         from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLFlashAttention2

verl/trainer/config.py

@@ -43,6 +43,7 @@ class DataConfig:
     rollout_batch_size: int = 512
     val_batch_size: int = -1
     format_prompt: Optional[str] = None
+    override_chat_template: Optional[str] = None
     shuffle: bool = True
     seed: int = 1
     max_pixels: int = 4194304
@@ -73,7 +74,7 @@ class AlgorithmConfig:
 
 @dataclass
 class TrainerConfig:
-    total_episodes: int = 10
+    total_epochs: int = 10
     max_steps: Optional[int] = None
     project_name: str = "easy_r1"
     experiment_name: str = "demo"

verl/trainer/main.py

@@ -20,7 +20,7 @@ from omegaconf import OmegaConf
 from ..single_controller.ray import RayWorkerGroup
 from ..utils.tokenizer import get_processor, get_tokenizer
 from ..workers.fsdp_workers import FSDPWorker
-from ..workers.reward import FunctionRewardManager
+from ..workers.reward import BatchFunctionRewardManager, SequentialFunctionRewardManager
 from .config import PPOConfig
 from .data_loader import create_dataloader
 from .ray_trainer import RayPPOTrainer, ResourcePoolManager, Role
@@ -38,11 +38,13 @@ class Runner:
         # instantiate tokenizer
         tokenizer = get_tokenizer(
             config.worker.actor.model.model_path,
+            override_chat_template=config.data.override_chat_template,
             trust_remote_code=config.worker.actor.model.trust_remote_code,
             use_fast=True,
         )
         processor = get_processor(
             config.worker.actor.model.model_path,
+            override_chat_template=config.data.override_chat_template,
             trust_remote_code=config.worker.actor.model.trust_remote_code,
             use_fast=True,
         )
@@ -65,7 +67,14 @@ class Runner:
         }
         resource_pool_manager = ResourcePoolManager(resource_pool_spec=resource_pool_spec, mapping=mapping)
 
-        RemoteRewardManager = ray.remote(FunctionRewardManager).options(num_cpus=config.worker.reward.num_cpus)
+        if config.worker.reward.reward_type == "sequential":
+            RewardManager = SequentialFunctionRewardManager
+        elif config.worker.reward.reward_type == "batch":
+            RewardManager = BatchFunctionRewardManager
+        else:
+            raise NotImplementedError(f"Unknown reward type {config.worker.reward.reward_type}.")
+
+        RemoteRewardManager = ray.remote(RewardManager).options(num_cpus=config.worker.reward.num_cpus)
         reward_fn = RemoteRewardManager.remote(config.worker.reward, tokenizer)
         val_reward_fn = RemoteRewardManager.remote(config.worker.reward, tokenizer)
 
@@ -117,7 +126,6 @@ def main():
                         runtime_env=runtime_env)
         else:
             ray.init(runtime_env=runtime_env)
-
     runner = Runner.remote()
     ray.get(runner.run.remote(ppo_config))
 

verl/trainer/ray_trainer.py

@@ -247,7 +247,7 @@ class RayPPOTrainer:
         if config.trainer.max_steps is not None:
             self.training_steps = config.trainer.max_steps
         else:
-            self.training_steps = len(train_dataloader) * config.trainer.total_episodes
+            self.training_steps = len(train_dataloader) * config.trainer.total_epochs
 
         config.worker.actor.optim.training_steps = self.training_steps
         config.worker.critic.optim.training_steps = self.training_steps
@@ -473,7 +473,7 @@ class RayPPOTrainer:
             if self.config.trainer.val_only:
                 return
 
-        for _ in tqdm(range(self.config.trainer.total_episodes), desc="Episode", position=0):
+        for _ in tqdm(range(self.config.trainer.total_epochs), desc="Epoch", position=0):
             for batch_dict in tqdm(self.train_dataloader, desc="Running step", position=1):
                 self.global_step += 1
                 if self.global_step > self.training_steps:

verl/utils/checkpoint/fsdp_checkpoint_manager.py

@@ -55,11 +55,13 @@ class FSDPCheckpointManager(BaseCheckpointManager):
         # every rank download its own checkpoint
         model_path = os.path.join(path, f"model_world_size_{self.world_size}_rank_{self.rank}.pt")
         optim_path = os.path.join(path, f"optim_world_size_{self.world_size}_rank_{self.rank}.pt")
-        extra_state_path = os.path.join(path, f"extra_state_world_size_{self.world_size}_rank_{self.rank}.pt")
-        print(f"[rank-{self.rank}]: Loading from {model_path} and {optim_path} and {extra_state_path}.")
+        extra_path = os.path.join(path, f"extra_state_world_size_{self.world_size}_rank_{self.rank}.pt")
+        print(f"[rank-{self.rank}]: Loading model from {os.path.abspath(model_path)}.")
+        print(f"[rank-{self.rank}]: Loading optimizer from {os.path.abspath(optim_path)}.")
+        print(f"[rank-{self.rank}]: Loading extra_state from {os.path.abspath(extra_path)}.")
         model_state_dict = torch.load(model_path, weights_only=False)
         optim_state_dict = torch.load(optim_path, weights_only=False)
-        extra_state_dict = torch.load(extra_state_path, weights_only=False)
+        extra_state_dict = torch.load(extra_path, weights_only=False)
 
         state_dict_options = StateDictOptions(cpu_offload=True)
         set_state_dict(
@@ -91,7 +93,7 @@ class FSDPCheckpointManager(BaseCheckpointManager):
         extra_path = os.path.join(path, f"extra_state_world_size_{self.world_size}_rank_{self.rank}.pt")
 
         print(f"[rank-{self.rank}]: Saving model to {os.path.abspath(model_path)}.")
-        print(f"[rank-{self.rank}]: Saving checkpoint to {os.path.abspath(model_path)}.")
+        print(f"[rank-{self.rank}]: Saving optimizer to {os.path.abspath(optim_path)}.")
         print(f"[rank-{self.rank}]: Saving extra_state to {os.path.abspath(extra_path)}.")
         torch.save(model_state_dict, model_path)
         torch.save(optim_state_dict, optim_path)