├── verl ├── version │ └── version ├── utils │ ├── reward_score │ │ ├── livecodebench │ │ │ ├── lcb_runner │ │ │ │ ├── __init__.py │ │ │ │ ├── utils │ │ │ │ │ ├── scenarios.py │ │ │ │ │ └── path_utils.py │ │ │ │ ├── prompts │ │ │ │ │ └── __init__.py │ │ │ │ ├── evaluation │ │ │ │ │ ├── __init__.py │ │ │ │ │ ├── compute_code_execution_metrics.py │ │ │ │ │ ├── old_results_check.py │ │ │ │ │ └── pass_k_utils.py │ │ │ │ ├── benchmarks │ │ │ │ │ ├── __init__.py │ │ │ │ │ └── code_execution.py │ │ │ │ └── runner │ │ │ │ │ ├── claude_runner.py │ │ │ │ │ ├── mistral_runner.py │ │ │ │ │ ├── cohere_runner.py │ │ │ │ │ └── claude3_runner.py │ │ │ └── __init__.py │ │ ├── deepscaler_math │ │ │ ├── globals.py │ │ │ ├── __init__.py │ │ │ └── utils │ │ │ │ └── __init__.py │ │ └── __init__.py │ ├── checkpoint │ │ └── __init__.py │ ├── logger │ │ ├── __init__.py │ │ └── aggregate_logger.py │ ├── megatron │ │ ├── __init__.py │ │ ├── memory.py │ │ ├── sequence_parallel.py │ │ └── pipeline_parallel.py │ ├── rendezvous │ │ └── __init__.py │ ├── debug │ │ ├── __init__.py │ │ └── performance.py │ ├── __init__.py │ ├── dataset │ │ ├── __init__.py │ │ └── README.md │ ├── config.py │ ├── logging_utils.py │ ├── distributed.py │ ├── import_utils.py │ ├── ray_utils.py │ └── py_functional.py ├── trainer │ ├── runtime_env.yaml │ ├── config │ │ ├── evaluation.yaml │ │ ├── sft_trainer.yaml │ │ └── generation.yaml │ ├── __init__.py │ └── ppo │ │ └── __init__.py ├── models │ ├── __init__.py │ ├── llama │ │ ├── __init__.py │ │ └── megatron │ │ │ ├── checkpoint_utils │ │ │ └── __init__.py │ │ │ ├── layers │ │ │ ├── __init__.py │ │ │ └── parallel_rmsnorm.py │ │ │ └── __init__.py │ ├── transformers │ │ └── __init__.py │ ├── weight_loader_registry.py │ └── README.md ├── workers │ ├── __init__.py │ ├── reward_model │ │ ├── __init__.py │ │ ├── megatron │ │ │ └── __init__.py │ │ └── base.py │ ├── rollout │ │ ├── naive │ │ │ └── __init__.py │ │ ├── vllm_rollout │ │ │ └── __init__.py │ │ ├── __init__.py │ │ └── base.py │ ├── reward_manager │ │ └── __init__.py │ ├── actor │ │ ├── __init__.py │ │ └── base.py │ ├── critic │ │ ├── __init__.py │ │ └── base.py │ └── sharding_manager │ │ ├── base.py │ │ └── __init__.py ├── third_party │ ├── __init__.py │ └── vllm │ │ ├── vllm_v_0_3_1 │ │ └── __init__.py │ │ ├── vllm_v_0_4_2 │ │ └── __init__.py │ │ ├── vllm_v_0_5_4 │ │ ├── __init__.py │ │ └── hf_weight_loader.py │ │ ├── vllm_v_0_6_3 │ │ ├── __init__.py │ │ ├── tokenizer.py │ │ └── hf_weight_loader.py │ │ └── __init__.py ├── single_controller │ ├── base │ │ ├── megatron │ │ │ ├── __init__.py │ │ │ ├── worker.py │ │ │ └── worker_group.py │ │ ├── register_center │ │ │ ├── __init__.py │ │ │ └── ray.py │ │ └── __init__.py │ ├── ray │ │ └── __init__.py │ └── __init__.py └── __init__.py ├── Notice.txt ├── assets ├── 7b_eval.jpg ├── 32b_eval.jpg ├── 32b_perf.jpg └── skywork-or1-math-7b-multi-stage.png ├── docs ├── _static │ └── logo.png ├── requirements-docs.txt ├── README.md ├── advance │ ├── placement.rst │ └── megatron_extension.rst ├── Makefile └── faq │ └── faq.rst ├── or1_data └── eval │ ├── aime24.parquet │ └── aime25.parquet ├── .style.yapf ├── scripts └── format.sh ├── tests ├── ray │ ├── detached_worker │ │ ├── run.sh │ │ ├── README.md │ │ └── client.py │ ├── test_check_worker_alive.py │ ├── test_rvdz.py │ ├── test_ray_local_envs.py │ └── check_worker_alive │ │ └── main.py ├── e2e │ ├── arithmetic_sequence │ │ ├── data │ │ │ ├── test.parquet │ │ │ ├── train.parquet │ │ │ └── create_dataset.py │ │ ├── model │ │ │ └── model.safetensors │ │ └── rl │ │ │ └── README.md │ ├── __init__.py │ ├── run_ray_trainer_rmpad.sh │ ├── envs │ │ ├── __init__.py │ │ └── digit_completion │ │ │ └── __init__.py │ ├── run_qwen_gsm8k_function_rm_grpo.sh │ ├── run_qwen_gsm8k_function_rm_remax.sh │ ├── run_ray_trainer.sh │ ├── check_results.py │ ├── run_deepseek_megatron.sh │ ├── run_qwen_gsm8k_function_rm.sh │ ├── run_qwen_gsm8k_function_rm_no_rmpad.sh │ └── run_qwen_gsm8k_model_rm.sh ├── __init__.py ├── sft │ ├── run_sft.sh │ ├── run_sft_sp_loss_match.sh │ ├── run_sft_qwen05_sp2_liger.sh │ └── run_sft_qwen05_peft.sh ├── sanity │ ├── test_import.py │ └── check_license.py ├── verl │ └── utils │ │ └── dataset │ │ ├── test_rm_dataset.py │ │ └── test_rl_dataset.py └── gpu_utility │ └── test_ops.py ├── requirements.txt ├── .readthedocs.yaml ├── examples ├── generation │ └── run_deepseek_v2_lite_math.sh ├── sft │ └── gsm8k │ │ ├── run_gemma_7b.sh │ │ ├── run_gemma_2b.sh │ │ ├── run_deepseek_6b7.sh │ │ ├── run_qwen_05_sp2.sh │ │ ├── run_qwen_05_sp2_liger.sh │ │ └── run_qwen_05_peft.sh ├── split_placement │ └── run_deepseek7b_llm.sh ├── ppo_trainer │ ├── run_deepseek_math_gsm8k_megatron.sh │ ├── run_deepseek_full_hh_rlhf.sh │ ├── run_gemma.sh │ ├── run_deepseek7b_llm.sh │ ├── run_deepseek_megatron.sh │ ├── run_deepseek7b_llm_sp2.sh │ ├── run_qwen2-7b.sh │ └── run_qwen2.5-32b.sh └── remax_trainer │ ├── run_qwen2.5-3b_seq_balance.sh │ └── run_qwen2.5-7b_seq_balance.sh ├── .github └── workflows │ ├── dataset.yml │ ├── sandbox.yml │ ├── sanity.yml │ ├── vllm.yml │ ├── ray_test.yml │ ├── e2e_digit_completion.yml │ ├── e2e_lora.yml │ ├── yapf_format.yml │ ├── e2e_gsm8k_megatron.yml │ ├── model.yml │ └── e2e_sft.yml ├── docker ├── Dockerfile.ngc.vllm └── Dockerfile.vemlp.vllm.te ├── setup.py ├── .gitignore └── or1_scripts └── eval ├── eval_7b.sh └── eval_32b.sh /verl/version/version: -------------------------------------------------------------------------------- 1 | 0.1 -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /Notice.txt: -------------------------------------------------------------------------------- 1 | Copyright 2023-2024 Bytedance Ltd. and/or its affiliates -------------------------------------------------------------------------------- /assets/7b_eval.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/assets/7b_eval.jpg -------------------------------------------------------------------------------- /assets/32b_eval.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/assets/32b_eval.jpg -------------------------------------------------------------------------------- /assets/32b_perf.jpg: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/assets/32b_perf.jpg -------------------------------------------------------------------------------- /docs/_static/logo.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/docs/_static/logo.png -------------------------------------------------------------------------------- /or1_data/eval/aime24.parquet: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/or1_data/eval/aime24.parquet -------------------------------------------------------------------------------- /or1_data/eval/aime25.parquet: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/or1_data/eval/aime25.parquet -------------------------------------------------------------------------------- /.style.yapf: -------------------------------------------------------------------------------- 1 | [style] 2 | based_on_style = google 3 | column_limit = 120 4 | indent_width = 4 5 | split_arguments_when_comma_terminated: true -------------------------------------------------------------------------------- /scripts/format.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | pip3 install --upgrade yapf 3 | yapf -ir -vv --style ./.style.yapf verl tests single_controller examples -------------------------------------------------------------------------------- /assets/skywork-or1-math-7b-multi-stage.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/assets/skywork-or1-math-7b-multi-stage.png -------------------------------------------------------------------------------- /tests/ray/detached_worker/run.sh: -------------------------------------------------------------------------------- 1 | #!/bin/bash 2 | ray start --head --port=6379 3 | python3 server.py 4 | python3 client.py 5 | ray stop --force -------------------------------------------------------------------------------- /tests/e2e/arithmetic_sequence/data/test.parquet: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/tests/e2e/arithmetic_sequence/data/test.parquet -------------------------------------------------------------------------------- /tests/e2e/arithmetic_sequence/data/train.parquet: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/tests/e2e/arithmetic_sequence/data/train.parquet -------------------------------------------------------------------------------- /tests/e2e/arithmetic_sequence/model/model.safetensors: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/SkyworkAI/Skywork-OR1/HEAD/tests/e2e/arithmetic_sequence/model/model.safetensors -------------------------------------------------------------------------------- /verl/trainer/runtime_env.yaml: -------------------------------------------------------------------------------- 1 | working_dir: ./ 2 | excludes: ["/.git/"] 3 | env_vars: 4 | TORCH_NCCL_AVOID_RECORD_STREAMS: "1" 5 | VLLM_ATTENTION_BACKEND: "XFORMERS" -------------------------------------------------------------------------------- /docs/requirements-docs.txt: -------------------------------------------------------------------------------- 1 | # markdown suport 2 | recommonmark 3 | # markdown table suport 4 | sphinx-markdown-tables 5 | 6 | # theme default rtd 7 | 8 | # crate-docs-theme 9 | sphinx-rtd-theme -------------------------------------------------------------------------------- /verl/trainer/config/evaluation.yaml: -------------------------------------------------------------------------------- 1 | data: 2 | path: /tmp/math_Qwen2-7B-Instruct.parquet 3 | prompt_key: prompt 4 | response_key: responses 5 | data_source_key: data_source 6 | reward_model_key: reward_model -------------------------------------------------------------------------------- /verl/utils/reward_score/deepscaler_math/globals.py: -------------------------------------------------------------------------------- 1 | """ 2 | Global variables for Deepscaler repo. 3 | """ 4 | 5 | # Reward function constants 6 | THOUGHT_DELIMITER_START = "" 7 | THOUGHT_DELIMITER_END = "" -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/__init__.py: -------------------------------------------------------------------------------- 1 | from verl.utils.reward_score.livecodebench.unit_test import lcb_compute_score, prepare_unit_test_data 2 | 3 | from verl.utils.reward_score.livecodebench.compute_score import compute_score 4 | -------------------------------------------------------------------------------- /verl/utils/reward_score/deepscaler_math/__init__.py: -------------------------------------------------------------------------------- 1 | """Import reward-related classes and types from the reward module.""" 2 | 3 | from .reward_types import RewardConfig, RewardFn, RewardInput, RewardOutput, RewardType 4 | 5 | __all__ = ['RewardFn', 'RewardInput', 'RewardOutput', 'RewardType'] 6 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/utils/scenarios.py: -------------------------------------------------------------------------------- 1 | from enum import Enum 2 | 3 | 4 | class Scenario(Enum): 5 | codegeneration = "codegeneration" 6 | selfrepair = "selfrepair" 7 | testoutputprediction = "testoutputprediction" 8 | codeexecution = "codeexecution" 9 | -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | accelerate 2 | codetiming 3 | datasets 4 | dill 5 | flash-attn 6 | hydra-core 7 | numpy 8 | pandas 9 | peft 10 | pyarrow>=15.0.0 11 | pybind11 12 | ray>=2.38 13 | tensordict<0.6 14 | transformers<4.48 15 | vllm==0.6.3 16 | wandb 17 | liger-kernel 18 | pylatexenc 19 | pyext 20 | anthropic 21 | math_verify -------------------------------------------------------------------------------- /tests/ray/detached_worker/README.md: -------------------------------------------------------------------------------- 1 | # Detached Worker 2 | ## How to run (Only on a single node) 3 | - Start a local ray cluster: 4 | ```bash 5 | ray start --head --port=6379 6 | ``` 7 | - Run the server 8 | ```bash 9 | python3 server.py 10 | ``` 11 | - On another terminal, Run the client 12 | ```bash 13 | python3 client.py 14 | ``` 15 | -------------------------------------------------------------------------------- /.readthedocs.yaml: -------------------------------------------------------------------------------- 1 | # Read the Docs configuration file 2 | # See https://docs.readthedocs.io/en/stable/config-file/v2.html for details 3 | 4 | version: 2 5 | 6 | build: 7 | os: ubuntu-22.04 8 | tools: 9 | python: "3.8" 10 | 11 | sphinx: 12 | configuration: docs/conf.py 13 | 14 | python: 15 | install: 16 | - requirements: docs/requirements-docs.txt -------------------------------------------------------------------------------- /docs/README.md: -------------------------------------------------------------------------------- 1 | # verl documents 2 | 3 | ## Build the docs 4 | 5 | ```bash 6 | # Install dependencies. 7 | pip install -r requirements-docs.txt 8 | 9 | # Build the docs. 10 | make clean 11 | make html 12 | ``` 13 | 14 | ## Open the docs with your browser 15 | 16 | ```bash 17 | python -m http.server -d _build/html/ 18 | ``` 19 | Launch your browser and open localhost:8000. -------------------------------------------------------------------------------- /verl/utils/reward_score/deepscaler_math/utils/__init__.py: -------------------------------------------------------------------------------- 1 | """ 2 | This module provides utility functions for grading mathematical answers and extracting answers from LaTeX formatted strings. 3 | """ 4 | 5 | from verl.utils.reward_score.deepscaler_math.utils.utils import ( 6 | extract_answer, 7 | grade_answer_sympy, 8 | grade_answer_mathd, 9 | ) 10 | 11 | __all__ = [ 12 | "extract_answer", 13 | "grade_answer_sympy", 14 | "grade_answer_mathd" 15 | ] 16 | -------------------------------------------------------------------------------- /docs/advance/placement.rst: -------------------------------------------------------------------------------- 1 | Ray API Design Tutorial 2 | ======================================= 3 | 4 | We provide a tutorial for our Ray API design, including: 5 | 6 | - Ray basic concepts 7 | - Resource Pool and RayWorkerGroup 8 | - Data Dispatch, Execution and Collection 9 | - Initialize the RayWorkerGroup and execute the distributed computation in the given Resource Pool 10 | 11 | See details in `tutorial.ipynb `_. -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/prompts/__init__.py: -------------------------------------------------------------------------------- 1 | from verl.utils.reward_score.livecodebench.lcb_runner.prompts.code_execution import format_prompt_execution, format_prompt_execution_cot 2 | from verl.utils.reward_score.livecodebench.lcb_runner.prompts.code_generation import format_prompt_generation 3 | from verl.utils.reward_score.livecodebench.lcb_runner.prompts.test_output_prediction import format_prompt_test_output 4 | from verl.utils.reward_score.livecodebench.lcb_runner.prompts.self_repair import format_prompt_self_repair 5 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/evaluation/__init__.py: -------------------------------------------------------------------------------- 1 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation.compute_code_generation_metrics import codegen_metrics 2 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation.compute_code_execution_metrics import code_execution_metrics 3 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation.compute_test_output_prediction_metrics import ( 4 | test_output_metrics, 5 | ) 6 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation.pass_k_utils import extract_instance_results 7 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/benchmarks/__init__.py: -------------------------------------------------------------------------------- 1 | from verl.utils.reward_score.livecodebench.lcb_runner.benchmarks.code_generation import ( 2 | CodeGenerationProblem, 3 | load_code_generation_dataset, 4 | load_code_generation_dataset_not_fast, 5 | ) 6 | from verl.utils.reward_score.livecodebench.lcb_runner.benchmarks.test_output_prediction import ( 7 | TestOutputPredictionProblem, 8 | load_test_prediction_dataset, 9 | ) 10 | from verl.utils.reward_score.livecodebench.lcb_runner.benchmarks.code_execution import ( 11 | CodeExecutionProblem, 12 | load_code_execution_dataset, 13 | ) 14 | -------------------------------------------------------------------------------- /tests/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. -------------------------------------------------------------------------------- /tests/e2e/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/models/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/trainer/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/utils/checkpoint/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. -------------------------------------------------------------------------------- /verl/workers/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /tests/e2e/run_ray_trainer_rmpad.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | set -e -x 4 | 5 | python3 tests/e2e/arithmetic_sequence/rl/main_trainer.py \ 6 | data.train_files=tests/e2e/arithmetic_sequence/data/train.parquet \ 7 | data.val_files=tests/e2e/arithmetic_sequence/data/test.parquet \ 8 | actor_rollout_ref.model.path=tests/e2e/arithmetic_sequence/model \ 9 | actor_rollout_ref.rollout.name=vllm \ 10 | actor_rollout_ref.rollout.tensor_model_parallel_size=1 \ 11 | actor_rollout_ref.model.tokenizer_path=tests/e2e/arithmetic_sequence/model \ 12 | critic.model.path=Qwen/Qwen2.5-0.5B \ 13 | critic.model.use_remove_padding=True \ 14 | trainer.total_epochs=1 -------------------------------------------------------------------------------- /verl/models/llama/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/third_party/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/trainer/ppo/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/utils/logger/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/utils/megatron/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/models/transformers/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/utils/rendezvous/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /examples/generation/run_deepseek_v2_lite_math.sh: -------------------------------------------------------------------------------- 1 | python3 -m verl.trainer.main_generation \ 2 | trainer.nnodes=1 \ 3 | trainer.n_gpus_per_node=8 \ 4 | data.path=~/data/rlhf/gsm8k/test.parquet \ 5 | data.prompt_key=prompt \ 6 | data.n_samples=1 \ 7 | data.output_path=~/data/rlhf/math/deepseek_v2_lite_gen_test.parquet \ 8 | model.path=deepseek-ai/deepseek-llm-7b-chat \ 9 | +model.trust_remote_code=True \ 10 | rollout.temperature=1.0 \ 11 | rollout.top_k=50 \ 12 | rollout.top_p=0.7 \ 13 | rollout.prompt_length=2048 \ 14 | rollout.response_length=1024 \ 15 | rollout.tensor_model_parallel_size=2 \ 16 | rollout.gpu_memory_utilization=0.8 17 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_3_1/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_4_2/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_5_4/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_6_3/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /docs/Makefile: -------------------------------------------------------------------------------- 1 | # Minimal makefile for Sphinx documentation 2 | # 3 | 4 | # You can set these variables from the command line. 5 | SPHINXOPTS = 6 | SPHINXBUILD = sphinx-build 7 | SPHINXPROJ = verl 8 | SOURCEDIR = . 9 | BUILDDIR = _build 10 | 11 | # Put it first so that "make" without argument is like "make help". 12 | help: 13 | @$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) 14 | 15 | .PHONY: help Makefile 16 | 17 | # Catch-all target: route all unknown targets to Sphinx using the new 18 | # "make mode" option. $(O) is meant as a shortcut for $(SPHINXOPTS). 19 | %: Makefile 20 | @$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O) 21 | -------------------------------------------------------------------------------- /verl/single_controller/base/megatron/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/models/llama/megatron/checkpoint_utils/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/single_controller/base/register_center/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | -------------------------------------------------------------------------------- /verl/utils/debug/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .performance import log_gpu_memory_usage -------------------------------------------------------------------------------- /verl/workers/reward_model/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .base import BasePPORewardModel 16 | -------------------------------------------------------------------------------- /verl/workers/rollout/naive/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .naive_rollout import NaiveRollout 16 | -------------------------------------------------------------------------------- /verl/workers/rollout/vllm_rollout/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .vllm_rollout import vLLMRollout -------------------------------------------------------------------------------- /verl/workers/reward_model/megatron/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .reward_model import MegatronRewardModel 16 | -------------------------------------------------------------------------------- /tests/e2e/envs/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .digit_completion import DigitCompletion 16 | 17 | __all__ = ['DigitCompletion'] -------------------------------------------------------------------------------- /verl/utils/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from . import tokenizer 16 | from .tokenizer import * 17 | 18 | __all__ = tokenizer.__all__ -------------------------------------------------------------------------------- /verl/single_controller/base/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .worker import Worker 16 | from .worker_group import WorkerGroup, ClassWithInitArgs, ResourcePool 17 | -------------------------------------------------------------------------------- /verl/utils/dataset/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .rl_dataset import RLHFDataset 16 | from .rm_dataset import RMDataset 17 | from .sft_dataset import SFTDataset 18 | -------------------------------------------------------------------------------- /verl/workers/reward_manager/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 PRIME team and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .naive import NaiveRewardManager 16 | from .prime import PrimeRewardManager 17 | from .yr_code import YRRewardManager -------------------------------------------------------------------------------- /verl/workers/actor/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .base import BasePPOActor 16 | from .dp_actor import DataParallelPPOActor 17 | 18 | __all__ = ["BasePPOActor", "DataParallelPPOActor"] 19 | -------------------------------------------------------------------------------- /verl/workers/critic/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .base import BasePPOCritic 16 | from .dp_critic import DataParallelPPOCritic 17 | 18 | __all__ = ["BasePPOCritic", "DataParallelPPOCritic"] 19 | -------------------------------------------------------------------------------- /tests/sft/run_sft.sh: -------------------------------------------------------------------------------- 1 | # Tested with 2 & 4 GPUs 2 | 3 | set -x 4 | 5 | torchrun --standalone --nnodes=1 --nproc_per_node=8 \ 6 | -m verl.trainer.fsdp_sft_trainer \ 7 | data.train_files=$HOME/data/gsm8k/train.parquet \ 8 | data.val_files=$HOME/data/gsm8k/test.parquet \ 9 | data.prompt_key=extra_info \ 10 | data.response_key=extra_info \ 11 | +data.prompt_dict_keys=['question'] \ 12 | +data.response_dict_keys=['answer'] \ 13 | data.micro_batch_size_per_gpu=32 \ 14 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 15 | trainer.default_local_dir=$HOME/ckpts/ \ 16 | trainer.project_name=qwen2.5-sft \ 17 | trainer.experiment_name=gsm8k-sft-gemma-2b-it \ 18 | trainer.total_training_steps=1 \ 19 | trainer.logger=['console'] \ 20 | trainer.default_hdfs_dir=null $@ 21 | 22 | rm -rf $HOME/ckpts/ -------------------------------------------------------------------------------- /verl/workers/rollout/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .base import BaseRollout 16 | from .naive import NaiveRollout 17 | from .hf_rollout import HFRollout 18 | 19 | __all__ = ["BaseRollout", "NaiveRollout", "HFRollout"] 20 | -------------------------------------------------------------------------------- /verl/single_controller/ray/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .base import RayResourcePool, RayClassWithInitArgs, RayWorkerGroup, create_colocated_worker_cls 16 | from .megatron import (MegatronRayWorkerGroup, DistRankInfo, DistGlobalInfo) -------------------------------------------------------------------------------- /verl/utils/dataset/README.md: -------------------------------------------------------------------------------- 1 | # Dataset Format 2 | ## RLHF dataset 3 | We combine all the data sources into a single parquet files. We directly organize the prompt into the chat format so that multi-turn chats can be easily incorporated. In the prompt, we may add instruction following texts to guide the model output the answers in a particular format so that we can extract the answers. 4 | 5 | Math problems 6 | ```json 7 | { 8 | "data_source": "openai/gsm8k", 9 | "prompt": [{"role": "user", "content": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May? Let's think step by step and output the final answer after \"####\""}], 10 | "ability": "math", 11 | "reward_model": { 12 | "style": "rule", 13 | "ground_truth": ["72"] 14 | }, 15 | } 16 | ``` 17 | -------------------------------------------------------------------------------- /tests/sanity/test_import.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | 16 | def test_import(): 17 | import verl 18 | print(verl.__version__) 19 | 20 | 21 | def test_single_controller_import(): 22 | import verl.single_controller 23 | print(verl.single_controller.__version__) 24 | -------------------------------------------------------------------------------- /verl/single_controller/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import os 16 | 17 | version_folder = os.path.dirname(os.path.join(os.path.abspath(__file__))) 18 | 19 | with open(os.path.join(os.path.join(version_folder, os.pardir), 'version/version')) as f: 20 | __version__ = f.read().strip() 21 | -------------------------------------------------------------------------------- /tests/sft/run_sft_sp_loss_match.sh: -------------------------------------------------------------------------------- 1 | # Tested with 2 & 4 GPUs 2 | 3 | set -x 4 | 5 | torchrun --standalone --nnodes=1 --nproc_per_node=8 \ 6 | tests/sft/test_sp_loss_match.py \ 7 | data.train_files=$HOME/data/gsm8k/train.parquet \ 8 | data.val_files=$HOME/data/gsm8k/test.parquet \ 9 | data.prompt_key=extra_info \ 10 | data.response_key=extra_info \ 11 | +data.prompt_dict_keys=['question'] \ 12 | +data.response_dict_keys=['answer'] \ 13 | data.micro_batch_size=32 \ 14 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 15 | ulysses_sequence_parallel_size=2 \ 16 | use_remove_padding=True \ 17 | trainer.default_local_dir=$HOME/ckpts/ \ 18 | trainer.project_name=qwen2.5-sft \ 19 | trainer.experiment_name=gsm8k-sft-gemma-2b-it \ 20 | trainer.total_training_steps=1 \ 21 | trainer.logger=['console'] \ 22 | trainer.default_hdfs_dir=null $@ 23 | 24 | rm -rf $HOME/ckpts/ 25 | -------------------------------------------------------------------------------- /verl/models/llama/megatron/layers/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .parallel_attention import ParallelLlamaAttention 16 | from .parallel_decoder import ParallelLlamaDecoderLayer, ParallelLlamaDecoderLayerRmPad 17 | from .parallel_mlp import ParallelLlamaMLP 18 | from .parallel_rmsnorm import ParallelLlamaRMSNorm 19 | -------------------------------------------------------------------------------- /verl/utils/config.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from typing import Dict 16 | 17 | from omegaconf import DictConfig 18 | 19 | 20 | def update_dict_with_config(dictionary: Dict, config: DictConfig): 21 | for key in dictionary: 22 | if hasattr(config, key): 23 | dictionary[key] = getattr(config, key) 24 | -------------------------------------------------------------------------------- /verl/utils/logging_utils.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import logging 16 | 17 | 18 | def set_basic_config(level): 19 | """ 20 | This function sets the global logging format and level. It will be called when import verl 21 | """ 22 | logging.basicConfig(format='%(levelname)s:%(asctime)s:%(message)s', level=level) 23 | -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_gemma_7b.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | if [ "$#" -lt 2 ]; then 4 | echo "Usage: run_gemma_7b.sh [other_configs...]" 5 | exit 1 6 | fi 7 | 8 | nproc_per_node=$1 9 | save_path=$2 10 | 11 | # Shift the arguments so $@ refers to the rest 12 | shift 2 13 | 14 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 15 | -m verl.trainer.fsdp_sft_trainer \ 16 | data.train_files=$HOME/data/gsm8k/train.parquet \ 17 | data.val_files=$HOME/data/gsm8k/test.parquet \ 18 | data.prompt_key=prompt \ 19 | data.response_key=answer \ 20 | data.micro_batch_size_per_gpu=4 \ 21 | model.partial_pretrain=google/gemma-1.1-7b-it \ 22 | trainer.default_local_dir=$save_path \ 23 | trainer.project_name=gsm8k-sft \ 24 | trainer.experiment_name=gsm8k-sft-gemma-1.1-7b-it \ 25 | trainer.total_epochs=4 \ 26 | trainer.logger=['console','wandb'] \ 27 | trainer.default_hdfs_dir=null $@ -------------------------------------------------------------------------------- /tests/e2e/envs/digit_completion/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .task import DigitCompletion, generate_ground_truth_response 16 | from .tokenizer import CharTokenizer 17 | 18 | from transformers import AutoTokenizer, LlamaConfig 19 | 20 | AutoTokenizer.register(LlamaConfig, CharTokenizer, exist_ok=True) 21 | 22 | __all__ = ['DigitCompletion', 'generate_ground_truth_response', 'CharTokenizer'] -------------------------------------------------------------------------------- /verl/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import os 16 | 17 | version_folder = os.path.dirname(os.path.join(os.path.abspath(__file__))) 18 | 19 | with open(os.path.join(version_folder, 'version/version')) as f: 20 | __version__ = f.read().strip() 21 | 22 | from .protocol import DataProto 23 | 24 | from .utils.logging_utils import set_basic_config 25 | import logging 26 | 27 | set_basic_config(level=logging.WARNING) 28 | -------------------------------------------------------------------------------- /docs/faq/faq.rst: -------------------------------------------------------------------------------- 1 | Frequently Asked Questions 2 | ==================================== 3 | 4 | Ray related 5 | ------------ 6 | 7 | How to add breakpoint for debugging with distributed Ray? 8 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 9 | 10 | Please checkout the official debugging guide from Ray: https://docs.ray.io/en/latest/ray-observability/ray-distributed-debugger.html 11 | 12 | 13 | Distributed training 14 | ------------------------ 15 | 16 | How to run multi-node post-training with Ray? 17 | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ 18 | 19 | You can start a ray cluster and submit a ray job, following the official guide from Ray: https://docs.ray.io/en/latest/ray-core/starting-ray.html 20 | 21 | If your cluster is managed by Slurm, please refer to the guide for deploying Ray on Slurm: https://docs.ray.io/en/latest/cluster/vms/user-guides/community/slurm.html 22 | -------------------------------------------------------------------------------- /verl/models/llama/megatron/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from .modeling_llama_megatron import ( 16 | # original model with megatron 17 | ParallelLlamaModel, 18 | ParallelLlamaForCausalLM, 19 | # rmpad with megatron 20 | ParallelLlamaForCausalLMRmPad, 21 | ParallelLlamaForValueRmPad, 22 | # rmpad with megatron and pipeline parallelism 23 | ParallelLlamaForCausalLMRmPadPP, 24 | ParallelLlamaForValueRmPadPP) 25 | -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_gemma_2b.sh: -------------------------------------------------------------------------------- 1 | # Tested with 2 & 4 GPUs 2 | 3 | set -x 4 | 5 | if [ "$#" -lt 2 ]; then 6 | echo "Usage: run_gemma_2b.sh [other_configs...]" 7 | exit 1 8 | fi 9 | 10 | nproc_per_node=$1 11 | save_path=$2 12 | 13 | # Shift the arguments so $@ refers to the rest 14 | shift 2 15 | 16 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 17 | -m verl.trainer.fsdp_sft_trainer \ 18 | data.train_files=$HOME/data/gsm8k/train.parquet \ 19 | data.val_files=$HOME/data/gsm8k/test.parquet \ 20 | data.prompt_key=extra_info \ 21 | data.response_key=extra_info \ 22 | +data.prompt_dict_keys=['question'] \ 23 | +data.response_dict_keys=['answer'] \ 24 | data.micro_batch_size_per_gpu=4 \ 25 | model.partial_pretrain=google/gemma-2b-it \ 26 | trainer.default_local_dir=$save_path \ 27 | trainer.project_name=gsm8k-sft \ 28 | trainer.experiment_name=gsm8k-sft-gemma-2b-it \ 29 | trainer.total_epochs=2 \ 30 | trainer.logger=['console','wandb'] \ 31 | trainer.default_hdfs_dir=null $@ -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_deepseek_6b7.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | if [ "$#" -lt 2 ]; then 4 | echo "Usage: run_deepseek_6b7.sh [other_configs...]" 5 | exit 1 6 | fi 7 | 8 | nproc_per_node=$1 9 | save_path=$2 10 | 11 | # Shift the arguments so $@ refers to the rest 12 | shift 2 13 | 14 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 15 | -m verl.trainer.fsdp_sft_trainer \ 16 | data.train_files=$HOME/data/gsm8k/train.parquet \ 17 | data.val_files=$HOME/data/gsm8k/test.parquet \ 18 | data.prompt_key=extra_info \ 19 | data.response_key=extra_info \ 20 | +data.prompt_dict_keys=['question'] \ 21 | +data.response_dict_keys=['answer'] \ 22 | data.micro_batch_size_per_gpu=4 \ 23 | model.partial_pretrain=deepseek-ai/deepseek-coder-6.7b-instruct \ 24 | trainer.default_local_dir=$save_path \ 25 | trainer.project_name=gsm8k-sft \ 26 | trainer.experiment_name=gsm8k-sft-deepseek-coder-6.7b-instruct \ 27 | trainer.total_epochs=4 \ 28 | trainer.logger=['console','wandb'] \ 29 | trainer.default_hdfs_dir=null $@ -------------------------------------------------------------------------------- /verl/single_controller/base/register_center/ray.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import ray 16 | 17 | 18 | @ray.remote 19 | class WorkerGroupRegisterCenter: 20 | 21 | def __init__(self, rank_zero_info): 22 | self.rank_zero_info = rank_zero_info 23 | 24 | def get_rank_zero_info(self): 25 | return self.rank_zero_info 26 | 27 | 28 | def create_worker_group_register_center(name, info): 29 | return WorkerGroupRegisterCenter.options(name=name).remote(info) 30 | -------------------------------------------------------------------------------- /.github/workflows/dataset.yml: -------------------------------------------------------------------------------- 1 | name: dataset 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/dataset.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/dataset.yml 18 | 19 | 20 | 21 | jobs: 22 | ray: 23 | runs-on: [self-hosted, gpu] 24 | steps: 25 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 26 | with: 27 | fetch-depth: 0 28 | - name: Install the current repository 29 | run: | 30 | pip install -e .[test] --user 31 | - name: Running dataset tests 32 | run: | 33 | [ ! -d "$HOME/verl-data" ] && git clone --depth 1 https://github.com/eric-haibin-lin/verl-data ~/verl-data 34 | pytest -s -x tests/verl 35 | - name: Running ray test using cupy (move it to L20 when dockerfile ready) 36 | run: | 37 | cd tests/ray 38 | pytest -s -x test_rvdz.py -------------------------------------------------------------------------------- /verl/workers/sharding_manager/base.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | Sharding manager to implement HybridEngine 16 | """ 17 | 18 | from verl import DataProto 19 | 20 | 21 | class BaseShardingManager: 22 | 23 | def __enter__(self): 24 | pass 25 | 26 | def __exit__(self, exc_type, exc_value, traceback): 27 | pass 28 | 29 | def preprocess_data(self, data: DataProto) -> DataProto: 30 | return data 31 | 32 | def postprocess_data(self, data: DataProto) -> DataProto: 33 | return data 34 | -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_qwen_05_sp2.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | if [ "$#" -lt 2 ]; then 4 | echo "Usage: run_qwen_05_sp2.sh [other_configs...]" 5 | exit 1 6 | fi 7 | 8 | nproc_per_node=$1 9 | save_path=$2 10 | 11 | # Shift the arguments so $@ refers to the rest 12 | shift 2 13 | 14 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 15 | -m verl.trainer.fsdp_sft_trainer \ 16 | data.train_files=$HOME/data/gsm8k/train.parquet \ 17 | data.val_files=$HOME/data/gsm8k/test.parquet \ 18 | data.prompt_key=extra_info \ 19 | data.response_key=extra_info \ 20 | optim.lr=1e-4 \ 21 | +data.prompt_dict_keys=['question'] \ 22 | +data.response_dict_keys=['answer'] \ 23 | data.micro_batch_size=4 \ 24 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 25 | trainer.default_local_dir=$save_path \ 26 | trainer.project_name=gsm8k-sft \ 27 | trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct-sp2 \ 28 | trainer.logger=['console'] \ 29 | trainer.total_training_steps=1 \ 30 | trainer.default_hdfs_dir=null $@ \ 31 | ulysses_sequence_parallel_size=2 \ 32 | use_remove_padding=true 33 | -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_qwen_05_sp2_liger.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | if [ "$#" -lt 2 ]; then 4 | echo "Usage: run_qwen_05_sp2.sh [other_configs...]" 5 | exit 1 6 | fi 7 | 8 | nproc_per_node=$1 9 | save_path=$2 10 | 11 | # Shift the arguments so $@ refers to the rest 12 | shift 2 13 | 14 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 15 | -m verl.trainer.fsdp_sft_trainer \ 16 | data.train_files=$HOME/data/gsm8k/train.parquet \ 17 | data.val_files=$HOME/data/gsm8k/test.parquet \ 18 | data.prompt_key=extra_info \ 19 | data.response_key=extra_info \ 20 | optim.lr=1e-4 \ 21 | +data.prompt_dict_keys=['question'] \ 22 | +data.response_dict_keys=['answer'] \ 23 | data.micro_batch_size=4 \ 24 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 25 | model.use_liger=True \ 26 | trainer.default_local_dir=$save_path \ 27 | trainer.project_name=gsm8k-sft \ 28 | trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct-sp2-liger \ 29 | trainer.logger=['console'] \ 30 | trainer.default_hdfs_dir=null $@ \ 31 | ulysses_sequence_parallel_size=2 \ 32 | use_remove_padding=true 33 | -------------------------------------------------------------------------------- /tests/sft/run_sft_qwen05_sp2_liger.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | if [ "$#" -lt 2 ]; then 4 | echo "Usage: run_sft_qwen05_sp2_liger.sh [other_configs...]" 5 | exit 1 6 | fi 7 | 8 | nproc_per_node=$1 9 | save_path=$2 10 | 11 | # Shift the arguments so $@ refers to the rest 12 | shift 2 13 | 14 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 15 | -m verl.trainer.fsdp_sft_trainer \ 16 | data.train_files=$HOME/data/gsm8k/train.parquet \ 17 | data.val_files=$HOME/data/gsm8k/test.parquet \ 18 | data.prompt_key=extra_info \ 19 | data.response_key=extra_info \ 20 | optim.lr=1e-4 \ 21 | +data.prompt_dict_keys=['question'] \ 22 | +data.response_dict_keys=['answer'] \ 23 | data.micro_batch_size=4 \ 24 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 25 | model.use_liger=True \ 26 | trainer.default_local_dir=$save_path \ 27 | trainer.project_name=gsm8k-sft \ 28 | trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct-sp2-liger \ 29 | trainer.logger=['console'] \ 30 | trainer.total_training_steps=1 \ 31 | trainer.default_hdfs_dir=null $@ \ 32 | ulysses_sequence_parallel_size=2 \ 33 | use_remove_padding=true -------------------------------------------------------------------------------- /verl/models/weight_loader_registry.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | 16 | def get_weight_loader(arch: str): 17 | from verl.models.llama.megatron.checkpoint_utils.llama_loader import load_state_dict_to_megatron_llama 18 | _MODEL_WEIGHT_MEGATRON_LOADER_REGISTRY = {'LlamaForCausalLM': load_state_dict_to_megatron_llama} 19 | 20 | if arch in _MODEL_WEIGHT_MEGATRON_LOADER_REGISTRY: 21 | return _MODEL_WEIGHT_MEGATRON_LOADER_REGISTRY[arch] 22 | raise ValueError(f"Model architectures {arch} are not supported for now. " 23 | f"Supported architectures: {_MODEL_WEIGHT_MEGATRON_LOADER_REGISTRY.keys()}") 24 | -------------------------------------------------------------------------------- /verl/utils/distributed.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """Utilities for distributed training.""" 15 | import os 16 | 17 | 18 | def initialize_global_process_group(timeout_second=36000): 19 | import torch.distributed 20 | from datetime import timedelta 21 | torch.distributed.init_process_group('nccl', timeout=timedelta(seconds=timeout_second)) 22 | local_rank = int(os.environ["LOCAL_RANK"]) 23 | rank = int(os.environ["RANK"]) 24 | world_size = int(os.environ["WORLD_SIZE"]) 25 | 26 | if torch.distributed.is_initialized(): 27 | torch.cuda.set_device(local_rank) 28 | return local_rank, rank, world_size 29 | -------------------------------------------------------------------------------- /verl/workers/critic/base.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | Base class for a critic 16 | """ 17 | from abc import ABC, abstractmethod 18 | 19 | import torch 20 | 21 | from verl import DataProto 22 | 23 | __all__ = ['BasePPOCritic'] 24 | 25 | 26 | class BasePPOCritic(ABC): 27 | 28 | def __init__(self, config): 29 | super().__init__() 30 | self.config = config 31 | 32 | @abstractmethod 33 | def compute_values(self, data: DataProto) -> torch.Tensor: 34 | """Compute values""" 35 | pass 36 | 37 | @abstractmethod 38 | def update_critic(self, data: DataProto): 39 | """Update the critic""" 40 | pass 41 | -------------------------------------------------------------------------------- /.github/workflows/sandbox.yml: -------------------------------------------------------------------------------- 1 | name: sandbox 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/sandbox.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/sandbox.yml 18 | 19 | jobs: 20 | sandbox: 21 | runs-on: [self-hosted, l20-0] 22 | env: 23 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 24 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 25 | NO_PROXY: "localhost,127.0.0.1" 26 | HF_HUB_ENABLE_HF_TRANSFER: 1 27 | container: 28 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 29 | options: --gpus all --shm-size=10g 30 | steps: 31 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 32 | with: 33 | fetch-depth: 0 34 | - name: Install the current repository 35 | run: | 36 | pip3 install hf_transfer 37 | pip3 install -e .[test] 38 | pip3 install vllm==0.5.4 39 | - name: Running sandbox tests on 8 L20 GPUs 40 | run: | 41 | cd tests/sandbox 42 | pytest -s -x . 43 | -------------------------------------------------------------------------------- /.github/workflows/sanity.yml: -------------------------------------------------------------------------------- 1 | name: sanity 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/sanity.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/sanity.yml 18 | 19 | jobs: 20 | sanity: 21 | runs-on: ubuntu-latest 22 | strategy: 23 | matrix: 24 | python-version: ["3.10"] 25 | steps: 26 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 27 | - name: Set up Python ${{ matrix.python-version }} 28 | uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 29 | with: 30 | python-version: ${{ matrix.python-version }} 31 | - name: Install the current repository 32 | run: | 33 | pip install -e .[test] 34 | - name: Run sanity test 35 | run: | 36 | pytest -s -x tests/sanity 37 | - name: Run untility test 38 | run: | 39 | pytest -s -x tests/utility 40 | - name: Run license test 41 | run: | 42 | python3 tests/sanity/check_license.py --directory . 43 | -------------------------------------------------------------------------------- /examples/sft/gsm8k/run_qwen_05_peft.sh: -------------------------------------------------------------------------------- 1 | # Tested with 2 & 4 GPUs 2 | 3 | set -x 4 | 5 | if [ "$#" -lt 2 ]; then 6 | echo "Usage: run_qwen_05_peft.sh [other_configs...]" 7 | exit 1 8 | fi 9 | 10 | nproc_per_node=$1 11 | save_path=$2 12 | 13 | # Shift the arguments so $@ refers to the rest 14 | shift 2 15 | 16 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 17 | -m verl.trainer.fsdp_sft_trainer \ 18 | data.train_files=$HOME/data/gsm8k/train.parquet \ 19 | data.val_files=$HOME/data/gsm8k/test.parquet \ 20 | data.prompt_key=extra_info \ 21 | data.response_key=extra_info \ 22 | optim.lr=1e-4 \ 23 | +data.prompt_dict_keys=['question'] \ 24 | +data.response_dict_keys=['answer'] \ 25 | data.micro_batch_size_per_gpu=4 \ 26 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 27 | trainer.default_local_dir=$save_path \ 28 | trainer.project_name=gsm8k-sft \ 29 | trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct \ 30 | trainer.logger=['console'] \ 31 | trainer.total_epochs=1 \ 32 | trainer.default_hdfs_dir=null $@ \ 33 | model.lora_rank=32\ 34 | model.lora_alpha=16 \ 35 | model.target_modules=all-linear 36 | 37 | # Or you can do this: 38 | # model.target_modules=[q_proj,v_proj] \ 39 | -------------------------------------------------------------------------------- /verl/workers/rollout/base.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from abc import ABC, abstractmethod 16 | from typing import Iterable, Union 17 | 18 | from verl import DataProto 19 | 20 | __all__ = ['BaseRollout'] 21 | 22 | 23 | class BaseRollout(ABC): 24 | 25 | def __init__(self): 26 | """ 27 | 28 | Args: 29 | dataloader: an Iterable of TensorDict that consistently generates prompts. Note that the dataloader 30 | should handle when the training stops. 31 | """ 32 | super().__init__() 33 | 34 | @abstractmethod 35 | def generate_sequences(self, prompts: DataProto) -> DataProto: 36 | """Generate sequences""" 37 | pass 38 | -------------------------------------------------------------------------------- /tests/sft/run_sft_qwen05_peft.sh: -------------------------------------------------------------------------------- 1 | # Tested with 2 & 4 GPUs 2 | 3 | set -x 4 | 5 | if [ "$#" -lt 2 ]; then 6 | echo "Usage: run_sft_qwen05_peft.sh [other_configs...]" 7 | exit 1 8 | fi 9 | 10 | nproc_per_node=$1 11 | save_path=$2 12 | 13 | # Shift the arguments so $@ refers to the rest 14 | shift 2 15 | 16 | torchrun --standalone --nnodes=1 --nproc_per_node=$nproc_per_node \ 17 | -m verl.trainer.fsdp_sft_trainer \ 18 | data.train_files=$HOME/data/gsm8k/train.parquet \ 19 | data.val_files=$HOME/data/gsm8k/test.parquet \ 20 | data.prompt_key=extra_info \ 21 | data.response_key=extra_info \ 22 | optim.lr=1e-4 \ 23 | +data.prompt_dict_keys=['question'] \ 24 | +data.response_dict_keys=['answer'] \ 25 | data.micro_batch_size_per_gpu=4 \ 26 | model.partial_pretrain=Qwen/Qwen2.5-0.5B-Instruct \ 27 | trainer.default_local_dir=$save_path \ 28 | trainer.project_name=gsm8k-sft \ 29 | trainer.experiment_name=gsm8k-sft-qwen-2.5-0.5b-instruct \ 30 | trainer.logger=['console'] \ 31 | trainer.total_training_steps=1 \ 32 | trainer.default_hdfs_dir=null $@ \ 33 | model.lora_rank=32\ 34 | model.lora_alpha=16 \ 35 | model.target_modules=all-linear 36 | 37 | # Or you can do this: 38 | # model.target_modules=[q_proj,v_proj] \ 39 | -------------------------------------------------------------------------------- /verl/workers/sharding_manager/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from verl.utils.import_utils import is_vllm_available, is_megatron_core_available 16 | 17 | from .base import BaseShardingManager 18 | from .fsdp_ulysses import FSDPUlyssesShardingManager 19 | 20 | AllGatherPPModel = None 21 | 22 | if is_megatron_core_available() and is_vllm_available(): 23 | from .megatron_vllm import AllGatherPPModel, MegatronVLLMShardingManager 24 | elif AllGatherPPModel is not None: 25 | pass 26 | else: 27 | AllGatherPPModel = None 28 | MegatronVLLMShardingManager = None 29 | 30 | if is_vllm_available(): 31 | from .fsdp_vllm import FSDPVLLMShardingManager 32 | else: 33 | FSDPVLLMShardingManager = None 34 | -------------------------------------------------------------------------------- /.github/workflows/vllm.yml: -------------------------------------------------------------------------------- 1 | name: vllm 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/vllm.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/vllm.yml 18 | 19 | jobs: 20 | vllm: 21 | runs-on: [self-hosted, l20-0] 22 | env: 23 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 24 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 25 | NO_PROXY: "localhost,127.0.0.1" 26 | HF_HUB_ENABLE_HF_TRANSFER: 1 27 | container: 28 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 29 | options: --gpus all --shm-size=10g 30 | steps: 31 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 32 | with: 33 | fetch-depth: 0 34 | - name: Install the current repository 35 | run: | 36 | pip3 install hf_transfer 37 | pip3 install -e .[test] 38 | pip3 install vllm==0.5.4 39 | - name: Running vllm tests on 8 L20 GPUs 40 | run: | 41 | cd tests/rollout 42 | torchrun --standalone --nnodes=1 --nproc_per_node=8 $(which pytest) -s test_vllm_hf_loader.py 43 | -------------------------------------------------------------------------------- /.github/workflows/ray_test.yml: -------------------------------------------------------------------------------- 1 | name: ray 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/ray_test.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/ray_test.yml 18 | 19 | 20 | 21 | jobs: 22 | ray: 23 | runs-on: [self-hosted, l20-0] 24 | env: 25 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 26 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 27 | NO_PROXY: "localhost,127.0.0.1" 28 | HF_HUB_ENABLE_HF_TRANSFER: 1 29 | container: 30 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 31 | options: --gpus all --shm-size=10g 32 | steps: 33 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 34 | with: 35 | fetch-depth: 0 36 | - name: Install the current repository 37 | run: | 38 | pip install hf_transfer 39 | pip install -e .[test] 40 | pip install --upgrade "ray>=2.40.0" 41 | - name: Running ray tests that need 8 GPUs 42 | run: | 43 | cd tests/ray 44 | pytest -s -x --ignore=test_check_worker_alive.py --ignore=test_rvdz.py . 45 | -------------------------------------------------------------------------------- /verl/utils/debug/performance.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import torch 16 | import torch.distributed as dist 17 | import logging 18 | 19 | 20 | def log_gpu_memory_usage(head: str, logger: logging.Logger = None, level=logging.DEBUG, rank: int = 0): 21 | if (not dist.is_initialized()) or (rank is None) or (dist.get_rank() == rank): 22 | memory_allocated = torch.cuda.memory_allocated() / 1024**3 23 | memory_reserved = torch.cuda.memory_reserved() / 1024**3 24 | 25 | message = f'{head}, memory allocated (GB): {memory_allocated}, memory reserved (GB): {memory_reserved}' 26 | 27 | if logger is None: 28 | print(message) 29 | else: 30 | logger.log(msg=message, level=level) 31 | -------------------------------------------------------------------------------- /.github/workflows/e2e_digit_completion.yml: -------------------------------------------------------------------------------- 1 | name: e2e_digit_completion 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/e2e_digit_completion.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/e2e_digit_completion.yml 18 | - "tests/e2e/*.sh" 19 | 20 | 21 | 22 | jobs: 23 | e2e_digit_completion: 24 | runs-on: [self-hosted, l20-0] 25 | env: 26 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 27 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 28 | NO_PROXY: "localhost,127.0.0.1" 29 | HF_HUB_ENABLE_HF_TRANSFER: 1 30 | container: 31 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 32 | options: --gpus all --shm-size=10g 33 | steps: 34 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 35 | with: 36 | fetch-depth: 0 37 | - name: Install the current repository 38 | run: | 39 | pip3 install hf_transfer 40 | pip3 install -e .[test] 41 | - name: Running digit completon e2e training tests on 8 L20 GPUs 42 | run: | 43 | ray stop --force 44 | bash tests/e2e/run_ray_trainer.sh 45 | -------------------------------------------------------------------------------- /.github/workflows/e2e_lora.yml: -------------------------------------------------------------------------------- 1 | name: e2e_lora 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/e2e_lora.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/e2e_lora.yml 18 | - "tests/e2e/*.sh" 19 | 20 | 21 | 22 | jobs: 23 | e2e_lora: 24 | runs-on: [self-hosted, l20-1] 25 | env: 26 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 27 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 28 | NO_PROXY: "localhost,127.0.0.1" 29 | HF_HUB_ENABLE_HF_TRANSFER: 1 30 | container: 31 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 32 | options: --gpus all --shm-size=10g 33 | steps: 34 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 35 | with: 36 | fetch-depth: 0 37 | - name: Install the current repository 38 | run: | 39 | pip3 install hf_transfer peft 40 | pip3 install -e .[test] 41 | - name: Prepare gsm8k dataset 42 | run: | 43 | ray stop --force 44 | python3 examples/data_preprocess/gsm8k.py 45 | - name: Running gsm8k e2e training tests with LoRA 46 | run: | 47 | ray stop --force 48 | bash tests/sft/run_sft_qwen05_peft.sh 8 $HOME/ckpts/ -------------------------------------------------------------------------------- /tests/e2e/arithmetic_sequence/rl/README.md: -------------------------------------------------------------------------------- 1 | # Digit completion 2 | 3 | This is an example of solving a digit completion problem. The problem is defined as below: 4 | 5 | The prompt is a sequence of numbers with fixed difference. The agent's goal is to complete the next N numbers. 6 | If the max number is reached, the next number should be modulo with max number. 7 | 8 | For example, 9 | - prompt = [1, 2, 3] 10 | - N = 5 11 | - max_number = 6 12 | 13 | The response should be [4, 5, 6, 7%6, 8%6] = [4, 5, 6, 0, 1]. 14 | 15 | # Environment definition 16 | 17 | The core definition of the task is defined in verl/envs/digit_completion/task.py 18 | 19 | It is highly recommended to take a look at it for better understanding. 20 | 21 | 22 | 23 | # Run experiments 24 | 25 | The users are required to specify the config path and config name (and the relative model config path to the current working directory) 26 | 27 | ```bash 28 | # cd examples/arithmetic_sequence/rl 29 | 30 | # Specify the config path and config name (current working dir) 31 | python3 -m verl.trainer.ppo.ray_megatron_train_synchronous --config-path=$(pwd)/config --config-name='ray_megatron' 32 | 33 | # The default relative path of model config is 'config/model_config', if you want to change it, you can rewrite it in ray_megatron.yaml or using: 34 | python3 -m verl.trainer.ppo.ray_megatron_train_synchronous --config-path=$(pwd)/config --config-name='ray_megatron' ++model.base_path=config/model_config 35 | 36 | ``` 37 | 38 | -------------------------------------------------------------------------------- /.github/workflows/yapf_format.yml: -------------------------------------------------------------------------------- 1 | name: yapf 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/yapf_format.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/yapf_format.yml 18 | 19 | jobs: 20 | yapf: 21 | runs-on: ubuntu-latest 22 | strategy: 23 | matrix: 24 | python-version: ["3.12"] 25 | steps: 26 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 27 | # - name: checkout 28 | # run: | 29 | # commits=${{ github.event.pull_request.commits }} 30 | # if [[ -n "$commits" ]]; then 31 | # # Prepare enough depth for diffs with main 32 | # git fetch --depth="$(( commits + 1 ))" 33 | # fi 34 | - name: Set up Python ${{ matrix.python-version }} 35 | uses: actions/setup-python@0b93645e9fea7318ecaed2b359559ac225c90a2b # v5.3.0 36 | with: 37 | python-version: ${{ matrix.python-version }} 38 | - name: Install dependencies 39 | run: | 40 | python -m pip install --upgrade pip 41 | pip install --upgrade yapf 42 | pip install toml==0.10.2 43 | - name: Running yapf 44 | run: | 45 | yapf -r -vv -d --style=./.style.yapf verl tests examples 46 | -------------------------------------------------------------------------------- /verl/trainer/config/sft_trainer.yaml: -------------------------------------------------------------------------------- 1 | data: 2 | train_batch_size: 256 3 | micro_batch_size: null # will be deprecated, use micro_batch_size_per_gpu 4 | micro_batch_size_per_gpu: 4 # this is also val batch size 5 | train_files: ~/data/gsm8k/train.parquet 6 | val_files: ~/data/gsm8k/test.parquet 7 | prompt_key: question 8 | response_key: answer 9 | max_length: 1024 10 | truncation: error 11 | balance_dp_token: False 12 | chat_template: null 13 | model: 14 | partial_pretrain: ~/models/gemma-1.1-7b-it 15 | fsdp_config: 16 | wrap_policy: 17 | min_num_params: 0 18 | cpu_offload: False 19 | offload_params: False 20 | external_lib: null 21 | enable_gradient_checkpointing: False 22 | trust_remote_code: False 23 | lora_rank: 0 # Set to positive value to enable LoRA (e.g., 32) 24 | lora_alpha: 16 # LoRA scaling factor 25 | target_modules: all-linear # Target modules for LoRA adaptation 26 | use_liger: False 27 | optim: 28 | lr: 1e-5 29 | betas: [0.9, 0.95] 30 | weight_decay: 0.01 31 | warmup_steps_ratio: 0.1 32 | clip_grad: 1.0 33 | ulysses_sequence_parallel_size: 1 34 | use_remove_padding: False 35 | trainer: 36 | default_local_dir: /tmp/sft_model 37 | default_hdfs_dir: hdfs://tmp/experiments/gsm8k/gemma-1.1-7b-it/ # change the hdfs path here 38 | resume_path: null 39 | project_name: gsm8k-sft 40 | experiment_name: test 41 | total_epochs: 4 42 | total_training_steps: null 43 | logger: ['console'] 44 | seed: 1 45 | 46 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/utils/path_utils.py: -------------------------------------------------------------------------------- 1 | import pathlib 2 | 3 | from verl.utils.reward_score.livecodebench.lcb_runner.lm_styles import LanguageModel, LMStyle 4 | from verl.utils.reward_score.livecodebench.lcb_runner.utils.scenarios import Scenario 5 | 6 | 7 | def ensure_dir(path: str, is_file=True): 8 | if is_file: 9 | pathlib.Path(path).parent.mkdir(parents=True, exist_ok=True) 10 | else: 11 | pathlib.Path(path).mkdir(parents=True, exist_ok=True) 12 | return 13 | 14 | 15 | def get_cache_path(model_repr:str, args) -> str: 16 | scenario: Scenario = args.scenario 17 | n = args.n 18 | temperature = args.temperature 19 | path = f"cache/{model_repr}/{scenario}_{n}_{temperature}.json" 20 | ensure_dir(path) 21 | return path 22 | 23 | 24 | def get_output_path(model_repr:str, args) -> str: 25 | scenario: Scenario = args.scenario 26 | n = args.n 27 | temperature = args.temperature 28 | cot_suffix = "_cot" if args.cot_code_execution else "" 29 | path = f"output/{model_repr}/{scenario}_{n}_{temperature}{cot_suffix}.json" 30 | ensure_dir(path) 31 | return path 32 | 33 | 34 | def get_eval_all_output_path(model_repr:str, args) -> str: 35 | scenario: Scenario = args.scenario 36 | n = args.n 37 | temperature = args.temperature 38 | cot_suffix = "_cot" if args.cot_code_execution else "" 39 | path = f"output/{model_repr}/{scenario}_{n}_{temperature}{cot_suffix}_eval_all.json" 40 | return path 41 | -------------------------------------------------------------------------------- /verl/utils/import_utils.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | Utilities to check if packages are available. 16 | We assume package availability won't change during runtime. 17 | """ 18 | 19 | from functools import cache 20 | from typing import List 21 | 22 | 23 | @cache 24 | def is_megatron_core_available(): 25 | try: 26 | from megatron.core import parallel_state as mpu 27 | return True 28 | except ImportError: 29 | return False 30 | 31 | 32 | @cache 33 | def is_vllm_available(): 34 | try: 35 | import vllm 36 | return True 37 | except ImportError: 38 | return False 39 | 40 | 41 | def import_external_libs(external_libs=None): 42 | if external_libs is None: 43 | return 44 | if not isinstance(external_libs, List): 45 | external_libs = [external_libs] 46 | import importlib 47 | for external_lib in external_libs: 48 | importlib.import_module(external_lib) 49 | -------------------------------------------------------------------------------- /verl/utils/logger/aggregate_logger.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | A Ray logger will receive logging info from different processes. 16 | """ 17 | import numbers 18 | from typing import Dict 19 | 20 | 21 | def concat_dict_to_str(dict: Dict, step): 22 | output = [f'step:{step}'] 23 | for k, v in dict.items(): 24 | if isinstance(v, numbers.Number): 25 | output.append(f'{k}:{v:.3f}') 26 | output_str = ' - '.join(output) 27 | return output_str 28 | 29 | 30 | class LocalLogger: 31 | 32 | def __init__(self, remote_logger=None, enable_wandb=False, print_to_console=False): 33 | self.print_to_console = print_to_console 34 | if print_to_console: 35 | print('Using LocalLogger is deprecated. The constructor API will change ') 36 | 37 | def flush(self): 38 | pass 39 | 40 | def log(self, data, step): 41 | if self.print_to_console: 42 | print(concat_dict_to_str(data, step=step), flush=True) -------------------------------------------------------------------------------- /verl/utils/ray_utils.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | Contains commonly used utilities for ray 16 | """ 17 | 18 | import ray 19 | 20 | import concurrent.futures 21 | 22 | 23 | def parallel_put(data_list, max_workers=None): 24 | 25 | def put_data(index, data): 26 | return index, ray.put(data) 27 | 28 | if max_workers is None: 29 | max_workers = min(len(data_list), 16) 30 | 31 | with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor: 32 | data_list_f = [executor.submit(put_data, i, data) for i, data in enumerate(data_list)] 33 | res_lst = [] 34 | for future in concurrent.futures.as_completed(data_list_f): 35 | res_lst.append(future.result()) 36 | 37 | # reorder based on index 38 | output = [None for _ in range(len(data_list))] 39 | for res in res_lst: 40 | index, data_ref = res 41 | output[index] = data_ref 42 | 43 | return output 44 | -------------------------------------------------------------------------------- /tests/verl/utils/dataset/test_rm_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | import os 15 | 16 | from transformers import AutoTokenizer 17 | from verl.utils import hf_tokenizer 18 | from verl.utils.dataset.rm_dataset import RMDataset 19 | 20 | 21 | def get_rm_data(): 22 | # prepare test dataset 23 | url = "https://github.com/eric-haibin-lin/verl-data/raw/refs/heads/main/full_hh_rlhf/rm/test.parquet" 24 | local_folder = os.path.expanduser('~/verl-data/full_hh_rlhf/rm/') 25 | local_path = os.path.join(local_folder, 'test.parquet') 26 | os.makedirs(local_folder, exist_ok=True) 27 | return local_path 28 | 29 | 30 | def test_rm_dataset(): 31 | tokenizer = hf_tokenizer("facebook/opt-1.3b") 32 | local_path = get_rm_data() 33 | dataset = RMDataset(parquet_files=local_path, tokenizer=tokenizer, max_length=512) 34 | data = dataset[0]['input_ids'] 35 | output = tokenizer.batch_decode(data) 36 | assert len(output) > 1 37 | assert type(output[0]) == str 38 | -------------------------------------------------------------------------------- /tests/sanity/check_license.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | license_head_bytedance = "Copyright 2024 Bytedance Ltd. and/or its affiliates" 16 | license_head2_prime = "Copyright 2024 PRIME team and/or its affiliates" 17 | 18 | from pathlib import Path 19 | from argparse import ArgumentParser 20 | 21 | if __name__ == '__main__': 22 | parser = ArgumentParser() 23 | parser.add_argument('--directory', '-d', required=True, type=str) 24 | args = parser.parse_args() 25 | directory_in_str = args.directory 26 | 27 | pathlist = Path(directory_in_str).glob('**/*.py') 28 | for path in pathlist: 29 | # because path is object not string 30 | path_in_str = str(path.absolute()) 31 | print(path_in_str) 32 | with open(path_in_str, 'r', encoding='utf-8') as f: 33 | file_content = f.read() 34 | 35 | assert license_head_bytedance in file_content or \ 36 | license_head2_prime in file_content, f'file {path_in_str} does not contain license' 37 | -------------------------------------------------------------------------------- /tests/e2e/run_qwen_gsm8k_function_rm_grpo.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | 5 | python3 -m verl.trainer.main_ppo \ 6 | data.train_files=$HOME/data/gsm8k/train.parquet \ 7 | data.val_files=$HOME/data/gsm8k/test.parquet \ 8 | data.train_batch_size=1024 \ 9 | data.val_batch_size=1312 \ 10 | data.max_prompt_length=512 \ 11 | data.max_response_length=512 \ 12 | actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B \ 13 | actor_rollout_ref.actor.optim.lr=1e-6 \ 14 | actor_rollout_ref.model.use_remove_padding=True \ 15 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 16 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 17 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 18 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 19 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 25 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 26 | algorithm.kl_ctrl.kl_coef=0.001 \ 27 | algorithm.adv_estimator=grpo \ 28 | trainer.critic_warmup=0 \ 29 | trainer.logger=['console'] \ 30 | trainer.project_name='verl_example_gsm8k' \ 31 | trainer.experiment_name='qwen_e2e_ci_function_rm' \ 32 | trainer.n_gpus_per_node=8 \ 33 | trainer.nnodes=1 \ 34 | trainer.save_freq=-1 \ 35 | trainer.total_training_steps=1 $@ 36 | -------------------------------------------------------------------------------- /tests/e2e/run_qwen_gsm8k_function_rm_remax.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | 5 | python3 -m verl.trainer.main_ppo \ 6 | data.train_files=$HOME/data/gsm8k/train.parquet \ 7 | data.val_files=$HOME/data/gsm8k/test.parquet \ 8 | data.train_batch_size=1024 \ 9 | data.val_batch_size=1312 \ 10 | data.max_prompt_length=512 \ 11 | data.max_response_length=512 \ 12 | actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B \ 13 | actor_rollout_ref.actor.optim.lr=1e-6 \ 14 | actor_rollout_ref.model.use_remove_padding=True \ 15 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 16 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 17 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 18 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 19 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 25 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 26 | algorithm.kl_ctrl.kl_coef=0.001 \ 27 | algorithm.adv_estimator=remax \ 28 | trainer.critic_warmup=0 \ 29 | trainer.logger=['console'] \ 30 | trainer.project_name='verl_example_gsm8k' \ 31 | trainer.experiment_name='qwen_e2e_ci_function_rm' \ 32 | trainer.n_gpus_per_node=8 \ 33 | trainer.nnodes=1 \ 34 | trainer.save_freq=-1 \ 35 | trainer.total_training_steps=1 $@ 36 | -------------------------------------------------------------------------------- /docker/Dockerfile.ngc.vllm: -------------------------------------------------------------------------------- 1 | FROM nvcr.io/nvidia/pytorch:24.05-py3 2 | 3 | # uninstall nv-pytorch fork 4 | RUN pip3 uninstall pytorch-quantization \ 5 | pytorch-triton \ 6 | torch \ 7 | torch-tensorrt \ 8 | torchvision \ 9 | xgboost transformer_engine flash_attn \ 10 | apex megatron-core -y 11 | 12 | RUN pip3 install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://download.pytorch.org/whl/cu124 13 | 14 | # make sure torch version is kept 15 | RUN pip3 install --no-cache-dir \ 16 | "torch==2.4.0" \ 17 | accelerate \ 18 | codetiming \ 19 | datasets \ 20 | dill \ 21 | hydra-core \ 22 | numpy \ 23 | pybind11 \ 24 | tensordict \ 25 | "transformers<=4.46.0" 26 | 27 | # ray is installed via vllm 28 | RUN pip3 install --no-cache-dir vllm==0.6.3 29 | 30 | # we choose flash-attn v2.7.0 or v2.7.2 which contain pre-built wheels 31 | RUN pip3 install --no-cache-dir --no-build-isolation flash-attn==2.7.0.post2 32 | 33 | # install apex, set MAX_JOBS to avoid OOMs 34 | RUN MAX_JOBS=4 pip3 install -v --disable-pip-version-check --no-cache-dir --no-build-isolation \ 35 | --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" \ 36 | git+https://github.com/NVIDIA/apex 37 | 38 | # install Transformer Engine, which requires FA 2.5.8 39 | RUN MAX_JOBS=4 NINJA_FLAGS="-j4" pip3 install flash-attn==2.5.8 --no-cache-dir --no-build-isolation 40 | RUN MAX_JOBS=4 NINJA_FLAGS="-j4" pip3 install git+https://github.com/NVIDIA/TransformerEngine.git@v1.7 41 | 42 | # Pin wandb to v0.18 since v0.19.1 is released with ImportError 43 | RUN pip3 install wandb==0.18.7 py-spy 44 | -------------------------------------------------------------------------------- /tests/e2e/run_ray_trainer.sh: -------------------------------------------------------------------------------- 1 | #!/usr/bin/env bash 2 | 3 | set -e -x 4 | 5 | OUTPUT_FILE="/tmp/output_ray_trainer.txt" 6 | 7 | export PATH=$PATH:~/.local/bin 8 | 9 | rm -rf $OUTPUT_FILE 10 | python3 tests/e2e/arithmetic_sequence/rl/main_trainer.py \ 11 | data.train_files=tests/e2e/arithmetic_sequence/data/train.parquet \ 12 | data.val_files=tests/e2e/arithmetic_sequence/data/test.parquet \ 13 | data.train_batch_size=800 \ 14 | data.val_batch_size=200 \ 15 | data.max_prompt_length=16 \ 16 | data.max_response_length=32 \ 17 | data.return_raw_input_ids=True \ 18 | actor_rollout_ref.model.path=tests/e2e/arithmetic_sequence/model \ 19 | actor_rollout_ref.model.external_lib=tests.e2e.envs.digit_completion \ 20 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=200 \ 21 | actor_rollout_ref.actor.entropy_coeff=0 \ 22 | actor_rollout_ref.actor.optim.lr=1e-4 \ 23 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=200 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=200 \ 25 | actor_rollout_ref.rollout.name=hf \ 26 | actor_rollout_ref.rollout.tensor_model_parallel_size=1 \ 27 | critic.ppo_micro_batch_size_per_gpu=200 \ 28 | critic.model.path=tests/e2e/arithmetic_sequence/model \ 29 | critic.optim.lr=1e-3 \ 30 | algorithm.kl_ctrl.kl_coef=0.005 \ 31 | trainer.total_epochs=200 \ 32 | trainer.experiment_name=arithmetic_sequences \ 33 | trainer.logger=['console'] \ 34 | trainer.n_gpus_per_node=1 \ 35 | trainer.test_freq=1 \ 36 | trainer.save_freq=110 | tee $OUTPUT_FILE; 37 | 38 | python3 tests/e2e/check_results.py --output_file=$OUTPUT_FILE 39 | rm -rf $OUTPUT_FILE 40 | -------------------------------------------------------------------------------- /.github/workflows/e2e_gsm8k_megatron.yml: -------------------------------------------------------------------------------- 1 | name: e2e_gsm8k_megatron 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/e2e_gsm8k_megatron.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/e2e_gsm8k_megatron.yml 18 | - "tests/e2e/*.sh" 19 | 20 | 21 | 22 | jobs: 23 | e2e_gsm8k_megatron: 24 | runs-on: [self-hosted, l20-0] 25 | env: 26 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 27 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 28 | NO_PROXY: "localhost,127.0.0.1" 29 | HF_HUB_ENABLE_HF_TRANSFER: 1 30 | container: 31 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 32 | options: --gpus all --shm-size=10g 33 | steps: 34 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 35 | with: 36 | fetch-depth: 0 37 | - name: Install the current repository 38 | run: | 39 | pip3 install hf_transfer 40 | pip3 install -e .[test] 41 | - name: Prepare gsm8k dataset 42 | run: | 43 | python3 examples/data_preprocess/gsm8k.py 44 | - name: Running gsm8k e2e training tests on 8 L20 GPUs with Megatron 45 | run: | 46 | ray stop --force 47 | [ ! -d "$HOME/Megatron-LM" ] && git clone -b core_v0.4.0_verl https://github.com/eric-haibin-lin/Megatron-LM $HOME/Megatron-LM 48 | export PYTHONPATH=$PYTHONPATH:$HOME/Megatron-LM 49 | bash tests/e2e/run_deepseek_megatron.sh -------------------------------------------------------------------------------- /verl/utils/megatron/memory.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import torch 16 | 17 | 18 | class MemoryBuffer: 19 | 20 | def __init__(self, numel, numel_padded, dtype): 21 | self.numel = numel 22 | self.numel_padded = numel_padded 23 | self.dtype = dtype 24 | self.data = torch.zeros(self.numel_padded, 25 | dtype=self.dtype, 26 | device=torch.cuda.current_device(), 27 | requires_grad=False) 28 | 29 | def zero(self): 30 | """Reset the buffer to zero.""" 31 | self.data.zero_() 32 | 33 | def get(self, shape, start_index): 34 | """Return a tensor with the input `shape` as a view into the 35 | 1-D data starting at `start_index`.""" 36 | end_index = start_index + shape.numel() 37 | assert end_index <= self.numel, \ 38 | 'requested tensor is out of the buffer range.' 39 | buffer_tensor = self.data[start_index:end_index] 40 | buffer_tensor = buffer_tensor.view(shape) 41 | return buffer_tensor 42 | -------------------------------------------------------------------------------- /verl/trainer/config/generation.yaml: -------------------------------------------------------------------------------- 1 | 2 | trainer: 3 | nnodes: 1 4 | n_gpus_per_node: 8 5 | 6 | data: 7 | path: ~/data/rlhf/math/test.parquet 8 | prompt_key: prompt 9 | response_key: responses 10 | data_source_key: data_source 11 | reward_model_key: reward_model 12 | n_samples: 1 13 | output_path: /opt/tiger/math_Qwen2-7B-Instruct.parquet 14 | batch_size: 2048 15 | 16 | model: 17 | path: ~/models/Qwen2-7B-Instruct 18 | external_lib: null 19 | 20 | rollout: 21 | name: vllm 22 | temperature: 0.6 23 | top_k: -1 # 0 for hf rollout, -1 for vllm rollout 24 | top_p: 1 25 | prompt_length: 2048 26 | response_length: 32768 27 | dtype: bfloat16 28 | gpu_memory_utilization: 0.9 29 | ignore_eos: False 30 | micro_batch_size: 256 31 | enforce_eager: True 32 | free_cache_engine: True 33 | load_format: dummy_dtensor 34 | tensor_model_parallel_size: 1 35 | max_num_batched_tokens: 8192 36 | max_num_seqs: 1024 37 | log_prob_micro_batch_size: 8 38 | log_prob_micro_batch_size_per_gpu: 1 39 | do_sample: True 40 | n: 1 41 | n_val: 1 42 | enable_chunked_prefill: True 43 | disable_log_stats: True 44 | 45 | actor: 46 | strategy: fsdp # This is for backward-compatibility 47 | ulysses_sequence_parallel_size: 1 # sp size 48 | fsdp_config: 49 | wrap_policy: 50 | min_num_params: 0 51 | param_offload: False 52 | grad_offload: False 53 | optimizer_offload: False 54 | fsdp_size: -1 55 | optim: 56 | lr: 1e-6 57 | lr_warmup_steps_ratio: 0. # the total steps will be injected during runtime 58 | min_lr_ratio: null # only useful for warmup with cosine 59 | warmup_style: constant # select from constant/cosine 60 | total_training_steps: -1 # must be override by program 61 | -------------------------------------------------------------------------------- /.github/workflows/model.yml: -------------------------------------------------------------------------------- 1 | name: model_rmpad 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/model.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/model.yml 18 | 19 | 20 | 21 | jobs: 22 | model_rmpad: 23 | runs-on: [self-hosted, l20-1] 24 | env: 25 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 26 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 27 | NO_PROXY: "localhost,127.0.0.1" 28 | HF_HUB_ENABLE_HF_TRANSFER: 1 29 | container: 30 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 31 | options: --gpus all --shm-size=10g 32 | steps: 33 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 34 | with: 35 | fetch-depth: 0 36 | - name: Install the current repository and upgrade to latest transformers/flash_attn 37 | run: | 38 | pip3 install -e .[test] 39 | pip3 install --upgrade transformers 40 | - name: Running rmpad model tests on 8 L20 GPUs + flash_attn 2.5.8 41 | run: | 42 | pytest -s tests/model/test_transformer.py 43 | - name: Running rmpad model tests on 8 L20 GPUs + latest flash_attn 44 | run: | 45 | pip3 install --upgrade flash_attn --no-build-isolation 46 | pytest -s tests/model/test_transformer.py 47 | - name: Running rmpad model tests on 8 L20 GPUs + latest flash_attn 48 | run: | 49 | pip3 install hf_transfer 50 | torchrun --nproc_per_node=8 tests/checkpoint/test_fsdp_ckpt.py 51 | -------------------------------------------------------------------------------- /tests/ray/test_check_worker_alive.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import time 16 | import os 17 | import subprocess 18 | 19 | 20 | def test(): 21 | wait_time = 10 22 | 23 | my_env = os.environ.copy() 24 | my_env["WAIT_TIME"] = str(wait_time) 25 | 26 | p = subprocess.Popen(["python3", "-u", "./check_worker_alive/main.py"], env=my_env, stdout=subprocess.PIPE) 27 | 28 | count = 0 29 | while b"foo started" not in p.stdout.read(): 30 | time.sleep(1) 31 | count += 1 32 | if count > 40: 33 | raise RuntimeError("timeout for start foo in check_worker_alive/main.py") 34 | 35 | print( 36 | time.time(), 37 | f"wait 1.5 wait time {wait_time*1.5} to let signal returned to process but still not exceed process wait time") 38 | time.sleep(wait_time * 1.5) 39 | print(time.time(), f"start checking") 40 | assert p.poll() is not None, f"process {p} still alive, expecting signal raised abort" 41 | assert p.returncode != 0, f"process {p} exit with code 0, expecting not-zero exit code" 42 | print(f"test passed") 43 | 44 | 45 | if __name__ == "__main__": 46 | test() 47 | -------------------------------------------------------------------------------- /verl/single_controller/base/megatron/worker.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import os 16 | from dataclasses import dataclass 17 | from verl.single_controller.base.worker import Worker, DistRankInfo, DistGlobalInfo 18 | 19 | 20 | class MegatronWorker(Worker): 21 | 22 | def __init__(self, cuda_visible_devices=None) -> None: 23 | super().__init__(cuda_visible_devices) 24 | 25 | def get_megatron_global_info(self): 26 | from megatron.core import parallel_state as mpu 27 | tp_size = mpu.get_tensor_model_parallel_world_size() 28 | dp_size = mpu.get_data_parallel_world_size() 29 | pp_size = mpu.get_pipeline_model_parallel_world_size() 30 | info = DistGlobalInfo(tp_size=tp_size, dp_size=dp_size, pp_size=pp_size) 31 | return info 32 | 33 | def get_megatron_rank_info(self): 34 | from megatron.core import parallel_state as mpu 35 | tp_rank = mpu.get_tensor_model_parallel_rank() 36 | dp_rank = mpu.get_data_parallel_rank() 37 | pp_rank = mpu.get_pipeline_model_parallel_rank() 38 | info = DistRankInfo(tp_rank=tp_rank, dp_rank=dp_rank, pp_rank=pp_rank) 39 | return info -------------------------------------------------------------------------------- /docs/advance/megatron_extension.rst: -------------------------------------------------------------------------------- 1 | Add models with the Megatron-LM backend 2 | ========================================= 3 | 4 | Model 5 | ----------- 6 | 7 | The most challenging aspect to use the Megatron-LM backend is implementing 8 | the models for training. Currently, we implement Llama model that 9 | support data parallelism, tensor parallelism, pipeline parallelism (also 10 | vPP) and sequence parallelism. We also implement remove padding (sequence packing) on Llama 11 | model, which can be found in `modeling_llama_megatron.py `_. 12 | 13 | To support other model, users are required to implement: 14 | 15 | 1. Implemnt a model similar to ``modeling_llama_megatron.py`` that satisfy the 16 | parallelism requirements of Megatron-LM. Then register your model in 17 | the `registry.py `_. 18 | 2. Checkpoint utils that can load full checkpoint (e.g. huggingface 19 | checkpoint) to partitioned models during the runtime. Then register 20 | your loader to ``weight_loader_registry`` in `weight_loader_registry.py `_. 21 | 3. Weight loader that synchronize the weight from Megatron to rollout 22 | (vLLM) model. Note that both the actor model and rollout model are 23 | partitioned during runtime. So, it's advisable to map the model name 24 | in actor model implementation. Otherwise, you may need an additional 25 | name mapping and even weight transformation. The weight loader implementation 26 | is in `megatron_weight_loaders.py `_. -------------------------------------------------------------------------------- /tests/ray/test_rvdz.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import ray 16 | 17 | 18 | @ray.remote 19 | class TestWorker: 20 | 21 | def __init__(self, rank, world_size, group_name): 22 | self.rank = rank 23 | self.world_size = world_size 24 | self.group_name = group_name 25 | self.communicator = None 26 | 27 | def init(self): 28 | from verl.utils.rendezvous.ray_backend import create_nccl_communicator_in_ray 29 | self.communicator = create_nccl_communicator_in_ray(self.rank, self.world_size, self.group_name) 30 | 31 | def test(self): 32 | if self.communicator is None: 33 | return None 34 | return self.communicator.rank_id() 35 | 36 | 37 | def test_rvdz(): 38 | ray.init() 39 | 40 | group_name = "test_group" 41 | world_size = 2 42 | 43 | workers = [TestWorker.options(num_gpus=1).remote(rank, world_size, group_name) for rank in range(world_size)] 44 | 45 | ray.get([worker.init.remote() for worker in workers]) 46 | 47 | ranks = ray.get([worker.test.remote() for worker in workers]) 48 | 49 | assert ranks == [0, 1], f"expecting [0, 1], got {ranks}" 50 | 51 | ray.shutdown() 52 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_6_3/tokenizer.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # Copyright 2023 The vLLM team. 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # Adapted from https://github.com/vllm-project/vllm/blob/main/vllm/transformers_utils/tokenizer_group/tokenizer_group.py 15 | 16 | from typing import Optional 17 | 18 | from transformers import PreTrainedTokenizer 19 | from vllm.transformers_utils.tokenizer_group import TokenizerGroup 20 | from vllm.utils import LRUCache 21 | 22 | 23 | class TokenizerGroup(TokenizerGroup): 24 | """A group of tokenizers that can be used for LoRA adapters.""" 25 | 26 | def __init__(self, tokenizer: PreTrainedTokenizer, enable_lora: bool, max_num_seqs: int, 27 | max_input_length: Optional[int]): 28 | self.enable_lora = enable_lora 29 | self.max_input_length = max_input_length 30 | self.tokenizer = tokenizer 31 | self.lora_tokenizers = LRUCache[PreTrainedTokenizer](capacity=max_num_seqs) if enable_lora else None 32 | 33 | # FIXME(sgm): for simplicity, we assign the special token here 34 | @property 35 | def pad_token_id(self): 36 | return self.tokenizer.pad_token_id 37 | 38 | @property 39 | def eos_token_id(self): 40 | return self.tokenizer.eos_token_id 41 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/runner/claude_runner.py: -------------------------------------------------------------------------------- 1 | import os 2 | from time import sleep 3 | 4 | try: 5 | from anthropic import Anthropic 6 | except ImportError as e: 7 | pass 8 | 9 | from verl.utils.reward_score.livecodebench.lcb_runner.runner.base_runner import BaseRunner 10 | 11 | 12 | class ClaudeRunner(BaseRunner): 13 | client = Anthropic(api_key=os.getenv("ANTHROPIC_KEY")) 14 | 15 | def __init__(self, args, model): 16 | super().__init__(args, model) 17 | self.client_kwargs: dict[str | str] = { 18 | "model": args.model, 19 | "temperature": args.temperature, 20 | "max_tokens_to_sample": args.max_tokens, 21 | "top_p": args.top_p, 22 | } 23 | 24 | def _run_single(self, prompt: str) -> list[str]: 25 | 26 | def __run_single(counter): 27 | try: 28 | response = self.client.completions.create( 29 | prompt=prompt, 30 | **self.client_kwargs, 31 | ) 32 | content = response.completion 33 | return content 34 | except Exception as e: 35 | print("Exception: ", repr(e), "Sleeping for 20 seconds...") 36 | sleep(20 * (11 - counter)) 37 | counter = counter - 1 38 | if counter == 0: 39 | print(f"Failed to run model for {prompt}!") 40 | print("Exception: ", repr(e)) 41 | raise e 42 | return __run_single(counter) 43 | 44 | outputs = [] 45 | try: 46 | for _ in range(self.args.n): 47 | outputs.append(__run_single(10)) 48 | except Exception as e: 49 | raise e 50 | 51 | return outputs 52 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/runner/mistral_runner.py: -------------------------------------------------------------------------------- 1 | import os 2 | from time import sleep 3 | 4 | try: 5 | from mistralai.client import MistralClient 6 | except ImportError as e: 7 | pass 8 | 9 | from verl.utils.reward_score.livecodebench.lcb_runner.runner.base_runner import BaseRunner 10 | 11 | 12 | class MistralRunner(BaseRunner): 13 | client = MistralClient( 14 | api_key=os.environ["MISTRAL_API_KEY"], 15 | ) 16 | 17 | def __init__(self, args, model): 18 | super().__init__(args, model) 19 | self.client_kwargs: dict[str | str] = { 20 | "model": args.model, 21 | "temperature": args.temperature, 22 | "max_tokens": args.max_tokens, 23 | "top_p": args.top_p, 24 | } 25 | 26 | def _run_single(self, prompt: list[dict[str, str]]) -> list[str]: 27 | 28 | def __run_single(counter): 29 | try: 30 | response = self.client.chat( 31 | messages=prompt, 32 | **self.client_kwargs, 33 | ) 34 | content = response.choices[0].message.content 35 | return content 36 | except Exception as e: 37 | print("Exception: ", repr(e), "Sleeping for 20 seconds...") 38 | sleep(20 * (11 - counter)) 39 | counter = counter - 1 40 | if counter == 0: 41 | print(f"Failed to run model for {prompt}!") 42 | print("Exception: ", repr(e)) 43 | raise e 44 | return __run_single(counter) 45 | 46 | outputs = [] 47 | try: 48 | for _ in range(self.args.n): 49 | outputs.append(__run_single(10)) 50 | except Exception as e: 51 | raise e 52 | 53 | return outputs 54 | -------------------------------------------------------------------------------- /tests/e2e/check_results.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import argparse 16 | 17 | import numpy as np 18 | 19 | 20 | def extract_reward_from_line(line): 21 | # TODO: this function needs error handling 22 | try: 23 | key_vals = line.split(' - ') 24 | for key_val in key_vals: 25 | key, val = key_val.split(':') 26 | if key == 'critic/rewards/mean': 27 | reward = float(val) 28 | return reward 29 | return -np.inf 30 | except Exception: 31 | return -np.inf 32 | 33 | 34 | if __name__ == '__main__': 35 | parser = argparse.ArgumentParser() 36 | parser.add_argument('--output_file', required=True, type=str) 37 | 38 | args = parser.parse_args() 39 | 40 | with open(args.output_file, 'r') as f: 41 | output = f.read().split('\n') 42 | 43 | best_reward = -np.inf 44 | for line in output: 45 | if line.startswith('step'): 46 | reward = extract_reward_from_line(line) 47 | if reward > best_reward: 48 | best_reward = reward 49 | 50 | print(f'Best reward is {best_reward}') 51 | assert best_reward > 0.2, f'Best reward must be greater than 0.2. best_reward: {best_reward}' 52 | print('Check passes') 53 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/runner/cohere_runner.py: -------------------------------------------------------------------------------- 1 | import os 2 | from time import sleep 3 | 4 | try: 5 | import cohere 6 | except ImportError as e: 7 | pass 8 | 9 | from verl.utils.reward_score.livecodebench.lcb_runner.runner.base_runner import BaseRunner 10 | 11 | 12 | class CohereRunner(BaseRunner): 13 | client = cohere.Client(os.getenv("COHERE_API_KEY")) 14 | 15 | def __init__(self, args, model): 16 | super().__init__(args, model) 17 | self.client_kwargs: dict[str | str] = { 18 | "model": args.model, 19 | "temperature": args.temperature, 20 | "max_tokens": args.max_tokens, 21 | "p": args.top_p, 22 | } 23 | 24 | def _run_single(self, prompt: tuple[dict[str,str], str]) -> list[str]: 25 | chat_history, message = prompt 26 | 27 | def __run_single(counter): 28 | try: 29 | response = self.client.chat( 30 | message=message, 31 | chat_history=chat_history, 32 | **self.client_kwargs, 33 | ) 34 | content = response.text 35 | return content 36 | except Exception as e: 37 | print("Exception: ", repr(e), "Sleeping for 20 seconds...") 38 | sleep(20 * (11 - counter)) 39 | counter = counter - 1 40 | if counter == 0: 41 | print(f"Failed to run model for {prompt}!") 42 | print("Exception: ", repr(e)) 43 | raise e 44 | return __run_single(counter) 45 | 46 | outputs = [] 47 | try: 48 | for _ in range(self.args.n): 49 | outputs.append(__run_single(10)) 50 | except Exception as e: 51 | raise e 52 | 53 | return outputs 54 | -------------------------------------------------------------------------------- /verl/models/README.md: -------------------------------------------------------------------------------- 1 | # Models 2 | Common modelzoo such as huggingface/transformers stuggles when using Pytorch native model parallelism. Following the design principle of vLLM, we keep a simple, parallelizable, highly-optimized with packed inputs in verl. 3 | ## Adding a New Huggingface Model 4 | ### Step 1: Copy the model file from HF to verl 5 | - Add a new file under verl/models/hf 6 | - Copy ONLY the model file from huggingface/transformers/models to verl/models/hf 7 | 8 | ### Step 2: Modify the model file to use packed inputs 9 | - Remove all the code related to inference (kv cache) 10 | - Modify the inputs to include only 11 | - input_ids (total_nnz,) 12 | - cu_seqlens (total_nnz + 1,) 13 | - max_seqlen_in_batch: int 14 | - Note that this requires using flash attention with causal mask. 15 | 16 | ### Step 2.5: Add tests 17 | - Add a test to compare this version and the huggingface version 18 | - Following the infrastructure and add tests to tests/models/hf 19 | 20 | ### Step 3: Add a function to apply tensor parallelism 21 | - Please follow 22 | - https://pytorch.org/docs/stable/distributed.tensor.parallel.html 23 | - https://pytorch.org/tutorials/intermediate/TP_tutorial.html 24 | - General comments 25 | - Tensor Parallelism in native Pytorch is NOT auto-parallelism. The way it works is to specify how model parameters and input/output reshards using configs. These configs are then registered as hooks to perform input/output resharding before/after model forward. 26 | 27 | ### Step 4: Add a function to apply data parallelism 28 | - Please use FSDP2 APIs 29 | - See demo here https://github.com/pytorch/torchtitan/blob/main/torchtitan/parallelisms/parallelize_llama.py#L413 30 | 31 | ### Step 5: Add a function to apply pipeline parallelism 32 | - Comes in Pytorch 2.4 33 | - Currently only in alpha in nightly version 34 | - Check torchtitan for more details 35 | 36 | -------------------------------------------------------------------------------- /examples/split_placement/run_deepseek7b_llm.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | python3 main_ppo_split.py \ 4 | data.train_files=$HOME/data/gsm8k/train.parquet \ 5 | data.val_files=$HOME/data/gsm8k/test.parquet \ 6 | data.train_batch_size=1024 \ 7 | data.val_batch_size=1312 \ 8 | data.max_prompt_length=512 \ 9 | data.max_response_length=512 \ 10 | actor_rollout_ref.model.path=deepseek-ai/deepseek-llm-7b-chat \ 11 | actor_rollout_ref.actor.optim.lr=1e-6 \ 12 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 13 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \ 14 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 15 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 16 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 17 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \ 18 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 19 | actor_rollout_ref.rollout.name=vllm \ 20 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 21 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=8 \ 22 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 23 | critic.optim.lr=1e-5 \ 24 | critic.model.path=deepseek-ai/deepseek-llm-7b-chat \ 25 | critic.model.enable_gradient_checkpointing=False \ 26 | critic.ppo_micro_batch_size_per_gpu=8 \ 27 | critic.model.fsdp_config.param_offload=False \ 28 | critic.model.fsdp_config.grad_offload=False \ 29 | critic.model.fsdp_config.optimizer_offload=False \ 30 | algorithm.kl_ctrl.kl_coef=0.001 \ 31 | trainer.critic_warmup=0 \ 32 | trainer.logger=['console','wandb'] \ 33 | trainer.project_name='verl_example_gsm8k' \ 34 | trainer.experiment_name='deepseek_llm_7b_function_rm' \ 35 | trainer.n_gpus_per_node=8 \ 36 | trainer.nnodes=1 \ 37 | trainer.save_freq=-1 \ 38 | trainer.total_epochs=15 $@ 39 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | # setup.py is the fallback installation script when pyproject.toml does not work 16 | from setuptools import setup, find_packages 17 | import os 18 | 19 | version_folder = os.path.dirname(os.path.join(os.path.abspath(__file__))) 20 | 21 | with open(os.path.join(version_folder, 'verl/version/version')) as f: 22 | __version__ = f.read().strip() 23 | 24 | 25 | with open('requirements.txt') as f: 26 | required = f.read().splitlines() 27 | install_requires = [item.strip() for item in required if item.strip()[0] != '#'] 28 | 29 | extras_require = { 30 | 'test': ['pytest', 'yapf'] 31 | } 32 | 33 | setup( 34 | name='verl', 35 | version=__version__, 36 | package_dir={'': '.'}, 37 | packages=find_packages(where='.'), 38 | url='https://github.com/volcengine/verl', 39 | license='Apache 2.0', 40 | author='Bytedance - Seed - MLSys', 41 | author_email='zhangchi.usc1992@bytedance.com, gmsheng@connect.hku.hk', 42 | description='verl: Volcano Engine Reinforcement Learning for LLM', 43 | install_requires=install_requires, 44 | extras_require=extras_require, 45 | package_data={'': ['version/*'], 46 | 'verl': ['trainer/config/*.yaml'],}, 47 | include_package_data=True, 48 | ) -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/runner/claude3_runner.py: -------------------------------------------------------------------------------- 1 | import os 2 | from time import sleep 3 | 4 | try: 5 | from anthropic import Anthropic 6 | except ImportError as e: 7 | pass 8 | 9 | from verl.utils.reward_score.livecodebench.lcb_runner.runner.base_runner import BaseRunner 10 | 11 | 12 | class Claude3Runner(BaseRunner): 13 | client = Anthropic(api_key=os.getenv("ANTHROPIC_KEY")) 14 | 15 | def __init__(self, args, model): 16 | super().__init__(args, model) 17 | self.client_kwargs: dict[str | str] = { 18 | "model": args.model, 19 | "temperature": args.temperature, 20 | "max_tokens": args.max_tokens, 21 | "top_p": args.top_p, 22 | } 23 | 24 | def _run_single(self, prompt: tuple[str, str]) -> list[str]: 25 | 26 | def __run_single(counter): 27 | try: 28 | response = self.client.messages.create( 29 | system=prompt[0], 30 | messages=prompt[1], 31 | **self.client_kwargs, 32 | ) 33 | content = "\n".join([x.text for x in response.content]) 34 | return content 35 | except Exception as e: 36 | print("Exception: ", repr(e), "Sleeping for 20 seconds...") 37 | sleep(20 * (11 - counter)) 38 | counter = counter - 1 39 | if counter == 0: 40 | print(f"Failed to run model for {prompt}!") 41 | print("Exception: ", repr(e)) 42 | raise e 43 | return __run_single(counter) 44 | 45 | outputs = [] 46 | try: 47 | for _ in range(self.args.n): 48 | outputs.append(__run_single(10)) 49 | except Exception as e: 50 | raise e 51 | 52 | return outputs 53 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_deepseek_math_gsm8k_megatron.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | gsm8k_train_path=$HOME/data/gsm8k/train.parquet 4 | gsm8k_test_path=$HOME/data/gsm8k/test.parquet 5 | math_train_path=$HOME/data/math/train.parquet 6 | math_test_path=$HOME/data/math/test.parquet 7 | 8 | train_files="['$gsm8k_train_path', '$math_train_path']" 9 | test_files="['$gsm8k_test_path', '$math_test_path']" 10 | 11 | python3 -m verl.trainer.main_ppo --config-path=./config --config-name='ppo_megatron_trainer'\ 12 | data.train_files="$train_files" \ 13 | data.val_files="$test_files" \ 14 | data.train_batch_size=1024 \ 15 | data.val_batch_size=6312 \ 16 | data.max_prompt_length=1024 \ 17 | data.max_response_length=512 \ 18 | actor_rollout_ref.model.path=deepseek-ai/deepseek-coder-6.7b-instruct \ 19 | actor_rollout_ref.actor.optim.lr=1e-6 \ 20 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 21 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 22 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \ 23 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 24 | actor_rollout_ref.rollout.name=vllm \ 25 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 26 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \ 27 | critic.optim.lr=1e-5 \ 28 | critic.model.path=deepseek-ai/deepseek-coder-6.7b-instruct \ 29 | critic.model.enable_gradient_checkpointing=False \ 30 | critic.ppo_micro_batch_size_per_gpu=4 \ 31 | algorithm.kl_ctrl.kl_coef=0.001 \ 32 | trainer.critic_warmup=0 \ 33 | trainer.logger=['console','wandb'] \ 34 | trainer.project_name='verl_megatron_math_gsm8k_examples' \ 35 | trainer.experiment_name='deepseek_llm_7b_function_rm' \ 36 | trainer.n_gpus_per_node=8 \ 37 | trainer.nnodes=1 \ 38 | trainer.save_freq=-1 \ 39 | trainer.test_freq=5 \ 40 | trainer.total_epochs=100 $@ 41 | -------------------------------------------------------------------------------- /verl/utils/py_functional.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | Contain small python utility functions 16 | """ 17 | 18 | from typing import Dict 19 | from types import SimpleNamespace 20 | 21 | 22 | def union_two_dict(dict1: Dict, dict2: Dict): 23 | """Union two dict. Will throw an error if there is an item not the same object with the same key. 24 | 25 | Args: 26 | dict1: 27 | dict2: 28 | 29 | Returns: 30 | 31 | """ 32 | for key, val in dict2.items(): 33 | if key in dict1: 34 | assert dict2[key] == dict1[key], \ 35 | f'{key} in meta_dict1 and meta_dict2 are not the same object' 36 | dict1[key] = val 37 | 38 | return dict1 39 | 40 | 41 | def append_to_dict(data: Dict, new_data: Dict): 42 | for key, val in new_data.items(): 43 | if key not in data: 44 | data[key] = [] 45 | data[key].append(val) 46 | 47 | 48 | class NestedNamespace(SimpleNamespace): 49 | 50 | def __init__(self, dictionary, **kwargs): 51 | super().__init__(**kwargs) 52 | for key, value in dictionary.items(): 53 | if isinstance(value, dict): 54 | self.__setattr__(key, NestedNamespace(value)) 55 | else: 56 | self.__setattr__(key, value) 57 | -------------------------------------------------------------------------------- /verl/workers/reward_model/base.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | The base class for reward model 16 | """ 17 | 18 | from abc import ABC, abstractmethod 19 | 20 | from verl import DataProto 21 | 22 | 23 | class BasePPORewardModel(ABC): 24 | 25 | def __init__(self, config): 26 | self.config = config 27 | 28 | @abstractmethod 29 | def compute_reward(self, data: DataProto) -> DataProto: 30 | """Computing reward given input_ids. The transformers should output a tensor with shape 31 | [batch_size, sequence_length], and the value at [EOS] mask should be gathered. 32 | 33 | Args: 34 | data: must contain keys "input_ids", "attention_mask" and "position_ids". 35 | - input_ids: [batch_size, sequence_length] 36 | - attention_mask: [batch_size, sequence_length] 37 | - position_ids: [batch_size, sequence_length] 38 | 39 | Returns: a data pass protocol containing "reward". Only the [EOS] position contains the reward. 40 | Other position should have zero reward. Note that this may change in the future if we use 41 | dense reward. So, we leave the interface for general case. 42 | - reward: [batch_size, sequence_length]. 43 | 44 | """ 45 | pass 46 | -------------------------------------------------------------------------------- /tests/gpu_utility/test_ops.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | 16 | def test_flash_attn_cross_entropy(): 17 | from verl.utils.torch_functional import logprobs_from_logits_naive 18 | 19 | from verl.utils.debug import log_gpu_memory_usage 20 | 21 | from flash_attn.ops.triton.cross_entropy import cross_entropy_loss 22 | 23 | import torch 24 | from torch import nn 25 | 26 | log_gpu_memory_usage('At start') 27 | 28 | hidden_states = torch.randn(size=(2048, 5120), device='cuda', requires_grad=True, dtype=torch.bfloat16) 29 | 30 | linear = nn.Linear(in_features=5120, out_features=155136, bias=False, device='cuda', dtype=torch.bfloat16) 31 | 32 | logits = linear(hidden_states) 33 | 34 | # logits = logits.float() 35 | labels = torch.randint(low=0, high=155136, size=(2048,), device='cuda') 36 | 37 | log_gpu_memory_usage('before computation') 38 | # output = checkpoint.checkpoint(logprobs_from_logits, logits, labels, use_reentrant=True) 39 | output = -cross_entropy_loss(logits, labels)[0] 40 | # output = logprobs_from_logits(logits, labels) 41 | log_gpu_memory_usage('After forward') 42 | output.sum().backward() 43 | log_gpu_memory_usage('After backward') 44 | 45 | groundtruth = logprobs_from_logits_naive(logits.float(), labels) 46 | 47 | torch.testing.assert_close(output, groundtruth) 48 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_6_3/hf_weight_loader.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # Copyright 2023 The vLLM team. 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/model_loader 15 | 16 | from typing import Dict 17 | 18 | import torch.nn as nn 19 | from vllm.model_executor.model_loader.utils import set_default_torch_dtype 20 | 21 | 22 | def update_hf_weight_loader(): 23 | print("no hf weight loader need to be updated") 24 | return 25 | 26 | 27 | def load_hf_weights(actor_weights: Dict, vllm_model: nn.Module): 28 | assert isinstance(actor_weights, Dict) 29 | with set_default_torch_dtype(next(vllm_model.parameters()).dtype): # TODO 30 | if vllm_model.config.tie_word_embeddings and "lm_head.weight" in actor_weights.keys(): 31 | del actor_weights["lm_head.weight"] 32 | vllm_model.load_weights(actor_weights.items()) 33 | for _, module in vllm_model.named_modules(): 34 | quant_method = getattr(module, "quant_method", None) 35 | if quant_method is not None: 36 | quant_method.process_weights_after_loading(module) 37 | # FIXME: Remove this after Mixtral is updated 38 | # to use quant_method. 39 | if hasattr(module, "process_weights_after_loading"): 40 | module.process_weights_after_loading() 41 | vllm_model = vllm_model.cuda() 42 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_deepseek_full_hh_rlhf.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | train_files=$HOME/data/full_hh_rlhf/rl/train.parquet 4 | test_files=$HOME/data/full_hh_rlhf/rl/train.parquet # no use 5 | 6 | python3 -m verl.trainer.main_ppo --config-path=./config --config-name='ppo_megatron_trainer'\ 7 | data.train_files="$train_files" \ 8 | data.val_files="$test_files" \ 9 | data.train_batch_size=512 \ 10 | data.val_batch_size=128 \ 11 | data.max_prompt_length=128 \ 12 | data.max_response_length=128 \ 13 | actor_rollout_ref.model.path=deepseek-ai/deepseek-llm-7b-chat \ 14 | actor_rollout_ref.actor.optim.lr=1e-6 \ 15 | actor_rollout_ref.actor.ppo_mini_batch_size=128 \ 16 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 17 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \ 18 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 19 | actor_rollout_ref.rollout.name=vllm \ 20 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 21 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \ 22 | actor_rollout_ref.ref.param_offload=False \ 23 | critic.optim.lr=1e-5 \ 24 | critic.model.path=deepseek-ai/deepseek-llm-7b-chat \ 25 | critic.model.enable_gradient_checkpointing=False \ 26 | critic.ppo_micro_batch_size_per_gpu=4 \ 27 | reward_model.enable=True \ 28 | reward_model.megatron.tensor_model_parallel_size=4 \ 29 | reward_model.model.path=deepseek-ai/deepseek-llm-7b-chat \ 30 | reward_model.micro_batch_size_per_gpu=4 \ 31 | reward_model.param_offload=False \ 32 | algorithm.kl_ctrl.kl_coef=0.001 \ 33 | trainer.critic_warmup=0 \ 34 | trainer.logger=['console','wandb'] \ 35 | trainer.project_name='verl_megatron_full_hh_rlhf_examples' \ 36 | trainer.experiment_name='deepseek_llm_7b_model_rm' \ 37 | trainer.n_gpus_per_node=8 \ 38 | trainer.nnodes=1 \ 39 | trainer.save_freq=-1 \ 40 | trainer.test_freq=5 \ 41 | trainer.total_epochs=100 $@ 42 | -------------------------------------------------------------------------------- /tests/e2e/run_deepseek_megatron.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | # the config file used: verl/trainer/main_ppo/config/ppo_megatron_trainer.yaml 4 | 5 | huggingface-cli download deepseek-ai/deepseek-coder-1.3b-instruct 6 | 7 | python3 -m verl.trainer.main_ppo --config-path=config \ 8 | --config-name='ppo_megatron_trainer.yaml'\ 9 | data.train_files=$HOME/data/gsm8k/train.parquet \ 10 | data.val_files=$HOME/data/gsm8k/test.parquet \ 11 | data.train_batch_size=1024 \ 12 | data.val_batch_size=1312 \ 13 | data.max_prompt_length=512 \ 14 | data.max_response_length=512 \ 15 | actor_rollout_ref.model.path=deepseek-ai/deepseek-coder-1.3b-instruct \ 16 | actor_rollout_ref.actor.optim.lr=2e-6 \ 17 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 18 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 19 | actor_rollout_ref.actor.megatron.tensor_model_parallel_size=2 \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 25 | actor_rollout_ref.ref.megatron.tensor_model_parallel_size=2 \ 26 | critic.optim.lr=2e-5 \ 27 | critic.model.path=deepseek-ai/deepseek-coder-1.3b-instruct \ 28 | critic.model.enable_gradient_checkpointing=False \ 29 | critic.ppo_micro_batch_size_per_gpu=4 \ 30 | critic.megatron.tensor_model_parallel_size=2 \ 31 | algorithm.kl_ctrl.kl_coef=0.001 \ 32 | trainer.critic_warmup=0 \ 33 | trainer.logger=['console'] \ 34 | trainer.project_name='verl_megatron_gsm8k_examples' \ 35 | trainer.experiment_name='deepseek_llm_1b3_function_rm' \ 36 | trainer.n_gpus_per_node=8 \ 37 | trainer.nnodes=1 \ 38 | trainer.save_freq=-1 \ 39 | trainer.test_freq=1 \ 40 | trainer.total_epochs=15 \ 41 | trainer.total_training_steps=3 $@ 42 | -------------------------------------------------------------------------------- /verl/third_party/vllm/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from importlib.metadata import version, PackageNotFoundError 16 | 17 | 18 | def get_version(pkg): 19 | try: 20 | return version(pkg) 21 | except PackageNotFoundError: 22 | return None 23 | 24 | 25 | package_name = 'vllm' 26 | package_version = get_version(package_name) 27 | 28 | if package_version == '0.3.1': 29 | vllm_version = '0.3.1' 30 | from .vllm_v_0_3_1.llm import LLM 31 | from .vllm_v_0_3_1.llm import LLMEngine 32 | from .vllm_v_0_3_1 import parallel_state 33 | elif package_version == '0.4.2': 34 | vllm_version = '0.4.2' 35 | from .vllm_v_0_4_2.llm import LLM 36 | from .vllm_v_0_4_2.llm import LLMEngine 37 | from .vllm_v_0_4_2 import parallel_state 38 | elif package_version == '0.5.4': 39 | vllm_version = '0.5.4' 40 | from .vllm_v_0_5_4.llm import LLM 41 | from .vllm_v_0_5_4.llm import LLMEngine 42 | from .vllm_v_0_5_4 import parallel_state 43 | elif package_version == '0.6.3': 44 | vllm_version = '0.6.3' 45 | from .vllm_v_0_6_3.llm import LLM 46 | from .vllm_v_0_6_3.llm import LLMEngine 47 | from .vllm_v_0_6_3 import parallel_state 48 | else: 49 | raise ValueError( 50 | f'vllm version {package_version} not supported. Currently supported versions are 0.3.1, 0.4.2, 0.5.4 and 0.6.3.' 51 | ) 52 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_gemma.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | python3 -m verl.trainer.main_ppo \ 4 | data.train_files=$HOME/data/gsm8k/train.parquet \ 5 | data.val_files=$HOME/data/gsm8k/test.parquet \ 6 | data.train_batch_size=512 \ 7 | data.val_batch_size=1312 \ 8 | data.max_prompt_length=1024 \ 9 | data.max_response_length=512 \ 10 | actor_rollout_ref.model.path=google/gemma-2-2b-it \ 11 | actor_rollout_ref.actor.optim.lr=1e-6 \ 12 | actor_rollout_ref.model.use_remove_padding=True \ 13 | actor_rollout_ref.actor.ppo_mini_batch_size=128 \ 14 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 15 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 16 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 17 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 18 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=4 \ 19 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 20 | actor_rollout_ref.rollout.name=vllm \ 21 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 22 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=4 \ 23 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 24 | critic.optim.lr=1e-5 \ 25 | critic.model.use_remove_padding=True \ 26 | critic.model.path=google/gemma-2-2b-it \ 27 | critic.model.enable_gradient_checkpointing=False \ 28 | critic.ppo_micro_batch_size_per_gpu=4 \ 29 | critic.model.fsdp_config.param_offload=False \ 30 | critic.model.fsdp_config.grad_offload=False \ 31 | critic.model.fsdp_config.optimizer_offload=False \ 32 | algorithm.kl_ctrl.kl_coef=0.001 \ 33 | trainer.critic_warmup=0 \ 34 | trainer.logger=['console','wandb'] \ 35 | trainer.project_name='verl_example' \ 36 | trainer.experiment_name='gemma2b_function_rm' \ 37 | trainer.n_gpus_per_node=2 \ 38 | trainer.nnodes=1 \ 39 | trainer.save_freq=-1 \ 40 | trainer.test_freq=10 \ 41 | trainer.total_epochs=15 $@ 42 | -------------------------------------------------------------------------------- /docker/Dockerfile.vemlp.vllm.te: -------------------------------------------------------------------------------- 1 | # docker buildx build --platform linux/x86_64 -t "verlai/verl:$TAG" -f docker/$FILE . 2 | 3 | # the one in docker.io is an alias for the one veturbo 4 | # FROM vemlp-cn-beijing.cr.volces.com/veturbo/pytorch:2.4-cu124 5 | FROM docker.io/haibinlin/verl:v0.0.5-th2.4.0-cu124-base 6 | 7 | # only config pip index with https://pypi.tuna.tsinghua.edu.cn/simple if needed 8 | # unset for now 9 | RUN pip3 config unset global.index-url 10 | 11 | # transformers 4.47.0 contains the following bug: 12 | # AttributeError: 'Gemma2Attention' object has no attribute '_flash_attn_uses_top_left_mask' 13 | RUN pip3 install --no-cache-dir \ 14 | torch==2.4.0 \ 15 | accelerate \ 16 | codetiming \ 17 | dill \ 18 | hydra-core \ 19 | numpy \ 20 | pybind11 \ 21 | tensordict \ 22 | "transformers <= 4.46.0" 23 | 24 | RUN pip3 install --no-cache-dir flash-attn==2.7.0.post2 --no-build-isolation 25 | 26 | # vllm depends on ray, and veRL does not support ray > 2.37 27 | RUN pip3 install --no-cache-dir vllm==0.6.3 ray==2.10 28 | 29 | # install apex 30 | RUN MAX_JOBS=4 pip3 install -v --disable-pip-version-check --no-cache-dir --no-build-isolation \ 31 | --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" \ 32 | git+https://github.com/NVIDIA/apex 33 | 34 | # install Transformer Engine 35 | # - flash-attn pinned to 2.5.3 by TransformerEngine, switch to eric-haibin-lin/TransformerEngine.git@v1.7.0 to relax version req 36 | # - install with: MAX_JOBS=1 NINJA_FLAGS="-j1" TE_BUILD_WITH_NINJA=0 to avoid OOM 37 | # - cudnn is required by TransformerEngine 38 | # RUN CUDNN_PATH=/opt/conda/lib/python3.11/site-packages/nvidia/cudnn \ 39 | # pip3 install git+https://github.com/eric-haibin-lin/TransformerEngine.git@v1.7.0 40 | RUN MAX_JOBS=1 NINJA_FLAGS="-j1" pip3 install flash-attn==2.5.3 --no-cache-dir --no-build-isolation 41 | RUN MAX_JOBS=1 NINJA_FLAGS="-j1" pip3 install git+https://github.com/NVIDIA/TransformerEngine.git@v1.7 42 | -------------------------------------------------------------------------------- /tests/e2e/arithmetic_sequence/data/create_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from tests.e2e.envs.digit_completion import DigitCompletion, generate_ground_truth_response 16 | from torch.utils import data 17 | import os 18 | 19 | if __name__ == '__main__': 20 | simple_task = DigitCompletion(max_number=9, max_diff=9, max_num_in_response=9) 21 | all_prompts = simple_task.get_all_prompts() 22 | 23 | # 21 * 6 * 4 24 | train_data, test_data = data.random_split(all_prompts, lengths=[0.8, 0.2]) 25 | train_data = list(train_data) 26 | test_data = list(test_data) 27 | 28 | train_data = [[{'role': 'user', 'content': str(item)}] \ 29 | for item in train_data] 30 | test_data = [[{'role': 'user', 'content': str(item)}] \ 31 | for item in test_data] 32 | 33 | print(f'Size of train: {len(train_data)}, size of test: {len(test_data)}') 34 | 35 | train_data = {'prompt': train_data} 36 | test_data = {'prompt': test_data} 37 | 38 | model_folder = os.path.join(os.path.dirname(os.path.abspath(__file__))) 39 | 40 | import pandas as pd 41 | 42 | train_data_frame = pd.DataFrame(train_data) 43 | test_data_frame = pd.DataFrame(test_data) 44 | 45 | train_data_frame.to_parquet(os.path.join(model_folder, 'train.parquet')) 46 | test_data_frame.to_parquet(os.path.join(model_folder, 'test.parquet')) 47 | -------------------------------------------------------------------------------- /examples/remax_trainer/run_qwen2.5-3b_seq_balance.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export HF_DATASETS_OFFLINE=1 4 | export TRANSFORMERS_OFFLINE=1 5 | 6 | export VLLM_ATTENTION_BACKEND=XFORMERS 7 | 8 | python3 -m verl.trainer.main_ppo \ 9 | algorithm.adv_estimator=remax \ 10 | data.train_files=$HOME/data/gsm8k/train.parquet \ 11 | data.val_files=$HOME/data/gsm8k/train.parquet \ 12 | data.train_batch_size=512 \ 13 | data.val_batch_size=1312 \ 14 | data.max_prompt_length=512 \ 15 | data.max_response_length=1024 \ 16 | actor_rollout_ref.model.path=Qwen/Qwen2.5-3B-Instruct \ 17 | actor_rollout_ref.actor.optim.lr=1e-6 \ 18 | actor_rollout_ref.model.use_remove_padding=True \ 19 | actor_rollout_ref.actor.ppo_mini_batch_size=128 \ 20 | actor_rollout_ref.actor.use_dynamic_bsz=True \ 21 | actor_rollout_ref.actor.ppo_max_token_len_per_gpu=30000 \ 22 | actor_rollout_ref.actor.use_kl_loss=True \ 23 | actor_rollout_ref.actor.kl_loss_coef=0.001 \ 24 | actor_rollout_ref.actor.kl_loss_type=low_var_kl \ 25 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 26 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 27 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 28 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 29 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 30 | actor_rollout_ref.rollout.name=vllm \ 31 | actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \ 32 | actor_rollout_ref.rollout.n=4 \ 33 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 34 | algorithm.kl_ctrl.kl_coef=0.001 \ 35 | trainer.critic_warmup=0 \ 36 | trainer.logger=['console','wandb'] \ 37 | trainer.project_name='verl_remax_example_gsm8k' \ 38 | trainer.experiment_name='qwen2.5_3b_function_rm_kl1e-3' \ 39 | +trainer.val_before_train=False \ 40 | trainer.n_gpus_per_node=8 \ 41 | trainer.nnodes=1 \ 42 | trainer.save_freq=-1 \ 43 | trainer.test_freq=5 \ 44 | trainer.total_epochs=5 $@ -------------------------------------------------------------------------------- /examples/remax_trainer/run_qwen2.5-7b_seq_balance.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export HF_DATASETS_OFFLINE=1 4 | export TRANSFORMERS_OFFLINE=1 5 | 6 | export VLLM_ATTENTION_BACKEND=XFORMERS 7 | 8 | python3 -m verl.trainer.main_ppo \ 9 | algorithm.adv_estimator=remax \ 10 | data.train_files=$HOME/data/gsm8k/train.parquet \ 11 | data.val_files=$HOME/data/gsm8k/train.parquet \ 12 | data.train_batch_size=1024 \ 13 | data.val_batch_size=1312 \ 14 | data.max_prompt_length=512 \ 15 | data.max_response_length=1024 \ 16 | actor_rollout_ref.model.path=Qwen/Qwen2.5-7B-Instruct \ 17 | actor_rollout_ref.actor.optim.lr=1e-6 \ 18 | actor_rollout_ref.model.use_remove_padding=True \ 19 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 20 | actor_rollout_ref.actor.use_dynamic_bsz=True \ 21 | actor_rollout_ref.actor.ppo_max_token_len_per_gpu=24000 \ 22 | actor_rollout_ref.actor.use_kl_loss=True \ 23 | actor_rollout_ref.actor.kl_loss_coef=0.001 \ 24 | actor_rollout_ref.actor.kl_loss_type=low_var_kl \ 25 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 26 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 27 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 28 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 29 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 30 | actor_rollout_ref.rollout.name=vllm \ 31 | actor_rollout_ref.rollout.gpu_memory_utilization=0.8 \ 32 | actor_rollout_ref.rollout.n=4 \ 33 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 34 | algorithm.kl_ctrl.kl_coef=0.001 \ 35 | trainer.critic_warmup=0 \ 36 | trainer.logger=['console','wandb'] \ 37 | trainer.project_name='verl_remax_example_gsm8k' \ 38 | trainer.experiment_name='qwen2.5_7b_function_rm_kl1e-3' \ 39 | +trainer.val_before_train=False \ 40 | trainer.n_gpus_per_node=8 \ 41 | trainer.nnodes=1 \ 42 | trainer.save_freq=-1 \ 43 | trainer.test_freq=5 \ 44 | trainer.total_epochs=10 $@ -------------------------------------------------------------------------------- /tests/e2e/run_qwen_gsm8k_function_rm.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | 5 | python3 -m verl.trainer.main_ppo \ 6 | data.train_files=$HOME/data/gsm8k/train.parquet \ 7 | data.val_files=$HOME/data/gsm8k/test.parquet \ 8 | data.train_batch_size=1024 \ 9 | data.val_batch_size=1312 \ 10 | data.max_prompt_length=512 \ 11 | data.max_response_length=512 \ 12 | actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B \ 13 | actor_rollout_ref.actor.optim.lr=1e-6 \ 14 | actor_rollout_ref.model.use_remove_padding=True \ 15 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 16 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 17 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 18 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 19 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 25 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 26 | critic.optim.lr=1e-5 \ 27 | critic.model.use_remove_padding=True \ 28 | critic.model.path=Qwen/Qwen2.5-0.5B \ 29 | critic.model.enable_gradient_checkpointing=False \ 30 | critic.ppo_micro_batch_size_per_gpu=4 \ 31 | critic.model.fsdp_config.param_offload=False \ 32 | critic.model.fsdp_config.grad_offload=False \ 33 | critic.model.fsdp_config.optimizer_offload=False \ 34 | algorithm.kl_ctrl.kl_coef=0.001 \ 35 | trainer.critic_warmup=0 \ 36 | trainer.logger=['console'] \ 37 | trainer.project_name='verl_example_gsm8k' \ 38 | trainer.experiment_name='qwen_e2e_ci_function_rm' \ 39 | trainer.n_gpus_per_node=8 \ 40 | trainer.nnodes=1 \ 41 | trainer.save_freq=1 \ 42 | trainer.total_training_steps=1 $@ 43 | -------------------------------------------------------------------------------- /verl/utils/megatron/sequence_parallel.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | 16 | import torch 17 | import torch.nn.functional as F 18 | from megatron.core import parallel_state as mpu 19 | 20 | 21 | def mark_parameter_as_sequence_parallel(parameter): 22 | setattr(parameter, 'sequence_parallel', True) 23 | 24 | 25 | def is_sequence_parallel_param(param): 26 | return hasattr(param, 'sequence_parallel') and param.sequence_parallel 27 | 28 | 29 | def pad_to_sequence_parallel(unpad_tokens: torch.Tensor): 30 | """pad the tokens such that the total length is a multiple of sp world size 31 | 32 | Args: 33 | unpad_tokens: (total_nnz, ...). Tokens after removing padding 34 | 35 | Returns: 36 | 37 | """ 38 | total_nnz = unpad_tokens.shape[0] 39 | sp_world_size = mpu.get_tensor_model_parallel_world_size() 40 | 41 | if total_nnz % sp_world_size == 0: 42 | pad_size = 0 43 | else: 44 | pad_size = sp_world_size - total_nnz % sp_world_size 45 | 46 | if pad_size > 0: 47 | if unpad_tokens.ndim == 1: 48 | unpad_tokens = F.pad(unpad_tokens, (0, pad_size)) 49 | elif unpad_tokens.ndim == 2: 50 | unpad_tokens = F.pad(unpad_tokens, (0, 0, 0, pad_size)) 51 | else: 52 | raise NotImplementedError(f'Padding dim {unpad_tokens.ndim()} is not supported') 53 | 54 | return unpad_tokens 55 | -------------------------------------------------------------------------------- /tests/e2e/run_qwen_gsm8k_function_rm_no_rmpad.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | 5 | python3 -m verl.trainer.main_ppo \ 6 | data.train_files=$HOME/data/gsm8k/train.parquet \ 7 | data.val_files=$HOME/data/gsm8k/test.parquet \ 8 | data.train_batch_size=1024 \ 9 | data.val_batch_size=1312 \ 10 | data.max_prompt_length=512 \ 11 | data.max_response_length=512 \ 12 | actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B \ 13 | actor_rollout_ref.actor.optim.lr=1e-6 \ 14 | actor_rollout_ref.model.use_remove_padding=False \ 15 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 16 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 17 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 18 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 19 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 25 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 26 | critic.optim.lr=1e-5 \ 27 | critic.model.use_remove_padding=False \ 28 | critic.model.path=Qwen/Qwen2.5-0.5B \ 29 | critic.model.enable_gradient_checkpointing=False \ 30 | critic.ppo_micro_batch_size_per_gpu=4 \ 31 | critic.model.fsdp_config.param_offload=False \ 32 | critic.model.fsdp_config.grad_offload=False \ 33 | critic.model.fsdp_config.optimizer_offload=False \ 34 | algorithm.kl_ctrl.kl_coef=0.001 \ 35 | trainer.critic_warmup=0 \ 36 | trainer.logger=['console'] \ 37 | +trainer.val_before_train=False \ 38 | trainer.project_name='verl_example_gsm8k' \ 39 | trainer.experiment_name='qwen_e2e_ci_function_rm' \ 40 | trainer.n_gpus_per_node=8 \ 41 | trainer.nnodes=1 \ 42 | trainer.save_freq=-1 \ 43 | trainer.total_training_steps=1 $@ 44 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/evaluation/compute_code_execution_metrics.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | from concurrent.futures import ProcessPoolExecutor 3 | import tqdm 4 | 5 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation.utils_execute import BASE_IMPORTS, check_correctness 6 | 7 | def evaluate_score(args) -> list[bool]: 8 | gs, (c, i, o) = args 9 | 10 | execution_results = [] 11 | for g in gs: 12 | if i in g: 13 | pass 14 | else: 15 | code_to_execute = f"{BASE_IMPORTS}\n{c}\nassert {o} == {g}" 16 | execution_results.append(check_correctness(code_to_execute, 3)) 17 | if len(execution_results) == 0: 18 | execution_results = [False] * len(gs) 19 | return execution_results 20 | 21 | def pass_at_k(n, c, k): 22 | if n - c < k: return 1.0 23 | return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1)) 24 | 25 | def code_execution_metrics( 26 | samples, 27 | generations, 28 | ): 29 | # execute the code 30 | references = [(doc["code"], doc["input"], doc["output"]) for doc in samples] 31 | with ProcessPoolExecutor() as executor: 32 | args_list = zip(generations, references) 33 | results = executor.map(evaluate_score, args_list) 34 | all_results = list(results) 35 | 36 | # serial version 37 | # all_results = [] 38 | # for i in range(len(generations)): 39 | # generation = generations[i] 40 | # result = evaluate_score([generation, references[i]]) 41 | # all_results.append(result) 42 | 43 | # compute pass@1 44 | pass_at_1s = [] 45 | for execution_result in all_results: 46 | c, n = execution_result.count(True), len(execution_result) 47 | pass_at_1s.append(pass_at_k(n, c, 1)) 48 | metrics = {"pass@1": sum(pass_at_1s) / len(pass_at_1s) * 100} 49 | 50 | results = {} 51 | for i, r in enumerate(all_results): 52 | r_new = [] 53 | for _r in r: 54 | r_new.append([_r]) 55 | results[i] = r_new 56 | return [metrics, results] 57 | -------------------------------------------------------------------------------- /tests/ray/test_ray_local_envs.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | e2e test verl.single_controller.ray 16 | """ 17 | import os 18 | import ray 19 | 20 | from verl.single_controller.ray.base import RayResourcePool, RayClassWithInitArgs, RayWorkerGroup 21 | from verl.single_controller.base.worker import Worker 22 | from verl.single_controller.base.decorator import register, Dispatch, collect_all_to_all, Execute 23 | 24 | 25 | @ray.remote 26 | class TestActor(Worker): 27 | 28 | def __init__(self) -> None: 29 | super().__init__() 30 | 31 | def getenv(self, key): 32 | val = os.getenv(key, f"{key} not set") 33 | return val 34 | 35 | 36 | def test_basics(): 37 | ray.init() 38 | 39 | # create 4 workers, each hold a GPU 40 | resource_pool = RayResourcePool([4], use_gpu=True) 41 | class_with_args = RayClassWithInitArgs(cls=TestActor) 42 | 43 | worker_group = RayWorkerGroup(resource_pool=resource_pool, 44 | ray_cls_with_init=class_with_args, 45 | name_prefix="worker_group_basic") 46 | 47 | output = worker_group.execute_all_sync("getenv", key="RAY_LOCAL_WORLD_SIZE") 48 | assert output == ["4", "4", "4", "4"] 49 | 50 | output = worker_group.execute_all_sync("getenv", key="RAY_LOCAL_RANK") 51 | assert set(output) == set(["0", "1", "2", "3"]) 52 | 53 | ray.shutdown() 54 | 55 | 56 | if __name__ == '__main__': 57 | test_basics() 58 | -------------------------------------------------------------------------------- /verl/models/llama/megatron/layers/parallel_rmsnorm.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import numbers 16 | import torch 17 | from megatron.core import ModelParallelConfig 18 | from torch import nn 19 | from transformers import LlamaConfig 20 | 21 | from apex.normalization.fused_layer_norm import fused_rms_norm_affine 22 | from verl.utils.megatron import sequence_parallel as sp_utils 23 | 24 | 25 | class ParallelLlamaRMSNorm(nn.Module): 26 | 27 | def __init__(self, config: LlamaConfig, megatron_config: ModelParallelConfig): 28 | """ 29 | LlamaRMSNorm is equivalent to T5LayerNorm 30 | """ 31 | super().__init__() 32 | if isinstance(config.hidden_size, numbers.Integral): 33 | normalized_shape = (config.hidden_size,) 34 | self.normalized_shape = torch.Size(normalized_shape) 35 | self.weight = nn.Parameter(torch.ones(self.normalized_shape)) 36 | self.variance_epsilon = config.rms_norm_eps 37 | 38 | if megatron_config.sequence_parallel: 39 | sp_utils.mark_parameter_as_sequence_parallel(self.weight) 40 | 41 | def forward(self, hidden_states): 42 | return fused_rms_norm_affine(input=hidden_states, 43 | weight=self.weight, 44 | normalized_shape=self.normalized_shape, 45 | eps=self.variance_epsilon, 46 | memory_efficient=True) -------------------------------------------------------------------------------- /examples/ppo_trainer/run_deepseek7b_llm.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | python3 -m verl.trainer.main_ppo \ 4 | data.train_files=$HOME/data/gsm8k/train.parquet \ 5 | data.val_files=$HOME/data/gsm8k/test.parquet \ 6 | data.train_batch_size=1024 \ 7 | data.val_batch_size=1312 \ 8 | data.max_prompt_length=512 \ 9 | data.max_response_length=512 \ 10 | actor_rollout_ref.model.path=deepseek-ai/deepseek-llm-7b-chat \ 11 | actor_rollout_ref.actor.optim.lr=1e-6 \ 12 | actor_rollout_ref.model.use_remove_padding=True \ 13 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 14 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \ 15 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 16 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 17 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 18 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 19 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \ 20 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 21 | actor_rollout_ref.rollout.name=vllm \ 22 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 23 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \ 24 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 25 | critic.optim.lr=1e-5 \ 26 | critic.model.use_remove_padding=True \ 27 | critic.model.path=deepseek-ai/deepseek-llm-7b-chat \ 28 | critic.model.enable_gradient_checkpointing=True \ 29 | critic.ppo_micro_batch_size_per_gpu=32 \ 30 | critic.model.fsdp_config.param_offload=False \ 31 | critic.model.fsdp_config.grad_offload=False \ 32 | critic.model.fsdp_config.optimizer_offload=False \ 33 | algorithm.kl_ctrl.kl_coef=0.001 \ 34 | trainer.critic_warmup=0 \ 35 | trainer.logger=['console','wandb'] \ 36 | trainer.project_name='verl_example_gsm8k' \ 37 | trainer.experiment_name='deepseek_llm_7b_function_rm' \ 38 | trainer.n_gpus_per_node=8 \ 39 | trainer.nnodes=1 \ 40 | trainer.save_freq=-1 \ 41 | trainer.test_freq=1 \ 42 | trainer.total_epochs=15 $@ 43 | -------------------------------------------------------------------------------- /.github/workflows/e2e_sft.yml: -------------------------------------------------------------------------------- 1 | name: e2e_sft 2 | 3 | on: 4 | # Trigger the workflow on push or pull request, 5 | # but only for the main branch 6 | push: 7 | branches: 8 | - main 9 | paths: 10 | - "**/*.py" 11 | - .github/workflows/e2e_sft.yml 12 | pull_request: 13 | branches: 14 | - main 15 | paths: 16 | - "**/*.py" 17 | - .github/workflows/e2e_sft.yml 18 | - "tests/e2e/*.sh" 19 | 20 | 21 | 22 | jobs: 23 | e2e_sft: 24 | runs-on: [self-hosted, l20-1] 25 | env: 26 | HTTP_PROXY: ${{ secrets.PROXY_HTTP }} 27 | HTTPS_PROXY: ${{ secrets.PROXY_HTTPS }} 28 | NO_PROXY: "localhost,127.0.0.1" 29 | HF_HUB_ENABLE_HF_TRANSFER: 1 30 | container: 31 | image: verlai/verl:vemlp-th2.4.0-cu124-vllm0.6.3-ray2.10-te1.7-v0.0.3 32 | options: --gpus all --shm-size=10g 33 | steps: 34 | - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2 35 | with: 36 | fetch-depth: 0 37 | - name: Install the current repository 38 | run: | 39 | pip3 install hf_transfer 40 | pip3 install -e .[test] 41 | - name: Prepare gsm8k dataset 42 | run: | 43 | ray stop --force 44 | python3 examples/data_preprocess/gsm8k.py 45 | - name: Running gsm8k e2e training tests on 8 L20 GPUs with rmpad using function rm 46 | run: | 47 | ray stop --force 48 | bash tests/sft/run_sft.sh 49 | - name: Running gsm8k e2e training tests on 8 L20 GPUs with sequence parallism 50 | run: | 51 | ray stop --force 52 | bash examples/sft/gsm8k/run_qwen_05_sp2.sh 8 $HOME/ckpts/ 53 | - name: Check loss difference between sequence parallel vs. default implementation 54 | run: | 55 | ray stop --force 56 | bash tests/sft/run_sft_sp_loss_match.sh 57 | - name: Running gsm8k e2e training tests on 8 L20 GPUs with sequence parallism and liger 58 | run: | 59 | ray stop --force 60 | bash tests/sft/run_sft_qwen05_sp2_liger.sh 8 $HOME/ckpts/ 61 | -------------------------------------------------------------------------------- /verl/third_party/vllm/vllm_v_0_5_4/hf_weight_loader.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # Copyright 2023 The vLLM team. 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # Adapted from https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/models 15 | 16 | from typing import Dict, Union, Optional, Iterable, Tuple 17 | 18 | import torch 19 | import torch.nn as nn 20 | 21 | from vllm.model_executor.model_loader.utils import set_default_torch_dtype 22 | from vllm.model_executor.model_loader.weight_utils import default_weight_loader 23 | 24 | 25 | def update_hf_weight_loader(): 26 | print('no hf weight loader need to be updated') 27 | return 28 | 29 | 30 | def load_hf_weights(actor_weights: Dict, vllm_model: nn.Module): 31 | assert isinstance(actor_weights, Dict) 32 | with set_default_torch_dtype(next(vllm_model.parameters()).dtype): # TODO 33 | if vllm_model.config.tie_word_embeddings and "lm_head.weight" in actor_weights.keys(): 34 | del actor_weights["lm_head.weight"] 35 | vllm_model.load_weights(actor_weights.items()) 36 | for _, module in vllm_model.named_modules(): 37 | quant_method = getattr(module, "quant_method", None) 38 | if quant_method is not None: 39 | quant_method.process_weights_after_loading(module) 40 | # FIXME: Remove this after Mixtral is updated 41 | # to use quant_method. 42 | if hasattr(module, "process_weights_after_loading"): 43 | module.process_weights_after_loading() 44 | vllm_model = vllm_model.cuda() 45 | -------------------------------------------------------------------------------- /verl/utils/reward_score/__init__.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | # from . import gsm8k, math, prime_math, prime_code 15 | 16 | 17 | def _default_compute_score(data_source, solution_str, ground_truth): 18 | # if data_source == 'openai/gsm8k': 19 | # from . import gsm8k 20 | # res = gsm8k.compute_score(solution_str, ground_truth) 21 | # elif data_source in ['lighteval/MATH', 'DigitalLearningGmbH/MATH-lighteval']: 22 | # from . import math 23 | # res = math.compute_score(solution_str, ground_truth) 24 | # elif data_source in [ 25 | # 'numina_aops_forum', 'numina_synthetic_math', 'numina_amc_aime', 'numina_synthetic_amc', 'numina_cn_k12', 26 | # 'numina_olympiads' 27 | # ]: 28 | # from . import prime_math 29 | # res = prime_math.compute_score(solution_str, ground_truth) 30 | # elif data_source in ['codecontests', 'apps', 'codeforces', 'taco']: 31 | # from . import prime_code 32 | # res = prime_code.compute_score(solution_str, ground_truth, continuous=True) 33 | # else: 34 | # raise NotImplementedError 35 | # from . import math 36 | # res = math.compute_score(solution_str, ground_truth) 37 | 38 | from verl.utils.reward_score.deepscaler_math.math_reward import deepscaler_reward_fn 39 | res = deepscaler_reward_fn(solution_str, ground_truth) 40 | 41 | if isinstance(res, (int, float, bool)): 42 | return float(res) 43 | else: 44 | return float(res[0]) 45 | -------------------------------------------------------------------------------- /tests/ray/check_worker_alive/main.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | import time 16 | import sys 17 | import os 18 | 19 | import ray 20 | 21 | from verl.single_controller.ray.base import RayResourcePool, RayClassWithInitArgs, RayWorkerGroup 22 | from verl.single_controller.base.worker import Worker 23 | from verl.single_controller.base.decorator import register, Dispatch 24 | 25 | 26 | @ray.remote 27 | class TestActor(Worker): 28 | 29 | def __init__(self) -> None: 30 | super().__init__() 31 | 32 | @register(dispatch_mode=Dispatch.ONE_TO_ALL, blocking=False) 33 | def foo(self, wait_time): 34 | time.sleep(wait_time) 35 | sys.exit(1) 36 | 37 | 38 | if __name__ == "__main__": 39 | wait_time = int(os.getenv("WAIT_TIME", "10")) 40 | 41 | ray.init() 42 | 43 | # test single-node-no-partition 44 | print(f"test single-node-no-partition") 45 | resource_pool = RayResourcePool([2], use_gpu=True) 46 | class_with_args = RayClassWithInitArgs(cls=TestActor) 47 | 48 | print("create worker group") 49 | wg = RayWorkerGroup(resource_pool, class_with_args, name_prefix="test") 50 | 51 | wg.start_worker_aliveness_check(1) 52 | time.sleep(1) 53 | 54 | print(time.time(), "start foo") 55 | 56 | _ = wg.foo(wait_time) 57 | print("foo started") 58 | 59 | print(time.time(), 60 | f"wait 6x wait time {wait_time*6} to let signal returned to process but still not exceed process wait time") 61 | time.sleep(wait_time * 6) 62 | 63 | ray.shutdown() 64 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | **/*.pt 2 | **/checkpoints 3 | **/wget-log 4 | **/_build/ 5 | **/*.ckpt 6 | **/outputs 7 | **/*.tar.gz 8 | **/playground 9 | **/wandb 10 | 11 | # Byte-compiled / optimized / DLL files 12 | __pycache__/ 13 | *.py[cod] 14 | *$py.class 15 | dataset/* 16 | tensorflow/my_graph/* 17 | .idea/ 18 | # C extensions 19 | *.so 20 | 21 | # Distribution / packaging 22 | .Python 23 | env/ 24 | build/ 25 | develop-eggs/ 26 | dist/ 27 | downloads/ 28 | eggs/ 29 | .eggs/ 30 | lib/ 31 | lib64/ 32 | parts/ 33 | sdist/ 34 | var/ 35 | *.egg-info/ 36 | .installed.cfg 37 | *.egg 38 | 39 | # PyInstaller 40 | # Usually these files are written by a python script from a template 41 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 42 | *.manifest 43 | *.spec 44 | 45 | # Installer logs 46 | pip-log.txt 47 | pip-delete-this-directory.txt 48 | 49 | # Unit test / coverage reports 50 | htmlcov/ 51 | .tox/ 52 | .coverage 53 | .coverage.* 54 | .cache 55 | nosetests.xml 56 | coverage.xml 57 | *,cover 58 | .hypothesis/ 59 | 60 | # Translations 61 | *.mo 62 | *.pot 63 | 64 | # Django stuff: 65 | *.log 66 | local_settings.py 67 | 68 | # Flask stuff: 69 | instance/ 70 | .webassets-cache 71 | 72 | # Scrapy stuff: 73 | .scrapy 74 | 75 | # Sphinx documentation 76 | docs/_build/ 77 | 78 | # PyBuilder 79 | target/ 80 | 81 | # IPython Notebook 82 | .ipynb_checkpoints 83 | 84 | # pyenv 85 | .python-version 86 | 87 | # celery beat schedule file 88 | celerybeat-schedule 89 | 90 | # dotenv 91 | .env 92 | 93 | # virtualenv 94 | venv/ 95 | ENV/ 96 | 97 | # Spyder project settings 98 | .spyderproject 99 | 100 | # Rope project settings 101 | .ropeproject 102 | 103 | # vscode 104 | .vscode 105 | 106 | # Mac 107 | .DS_Store 108 | 109 | # output logs 110 | tests/e2e/toy_examples/deepspeed/synchronous/output.txt 111 | 112 | # vim 113 | *.swp 114 | 115 | # ckpt 116 | *.lock 117 | 118 | verl_ckpt/ 119 | yr_test/ 120 | tmp/ 121 | *.pkl 122 | *.jsonl 123 | *.json 124 | a.py 125 | b.py 126 | c.py 127 | yr_test.py 128 | or1_data/train/ 129 | temp 130 | or1_data/eval/livecodebench 131 | Skywork/ -------------------------------------------------------------------------------- /examples/ppo_trainer/run_deepseek_megatron.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | # prepare pre-trained model ckpt 4 | huggingface-cli download deepseek-ai/deepseek-llm-7b-chat --local-dir $HOME/models/deepseek-llm-7b-chat 5 | 6 | # ``actor_rollout_ref.rollout.tensor_model_parallel_size`` in theory could be different from 7 | # ``**.megatron.tensor_model_parallel_size`` 8 | 9 | # the config file used: verl/trainer/main_ppo/config/ppo_megatron_trainer.yaml 10 | 11 | python3 -m verl.trainer.main_ppo --config-path=config \ 12 | --config-name='ppo_megatron_trainer.yaml'\ 13 | data.train_files=$HOME/data/gsm8k/train.parquet \ 14 | data.val_files=$HOME/data/gsm8k/test.parquet \ 15 | data.train_batch_size=1024 \ 16 | data.val_batch_size=1312 \ 17 | data.max_prompt_length=512 \ 18 | data.max_response_length=512 \ 19 | actor_rollout_ref.model.path=$HOME/models/deepseek-llm-7b-chat \ 20 | actor_rollout_ref.actor.optim.lr=2e-6 \ 21 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 22 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 23 | actor_rollout_ref.actor.megatron.tensor_model_parallel_size=4 \ 24 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \ 25 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 26 | actor_rollout_ref.rollout.name=vllm \ 27 | actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \ 28 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 29 | actor_rollout_ref.ref.megatron.tensor_model_parallel_size=4 \ 30 | critic.optim.lr=2e-5 \ 31 | critic.model.path=$HOME/models/deepseek-llm-7b-chat \ 32 | critic.model.enable_gradient_checkpointing=False \ 33 | critic.ppo_micro_batch_size_per_gpu=4 \ 34 | critic.megatron.tensor_model_parallel_size=4 \ 35 | algorithm.kl_ctrl.kl_coef=0.001 \ 36 | trainer.critic_warmup=0 \ 37 | trainer.logger=['console','wandb'] \ 38 | trainer.project_name='verl_megatron_gsm8k_examples' \ 39 | trainer.experiment_name='deepseek_llm_7b_function_rm' \ 40 | trainer.n_gpus_per_node=8 \ 41 | trainer.nnodes=1 \ 42 | trainer.save_freq=-1 \ 43 | trainer.total_epochs=15 \ 44 | +trainer.val_before_train=False $@ 45 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_deepseek7b_llm_sp2.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | python3 -m verl.trainer.main_ppo \ 4 | data.train_files=$HOME/data/gsm8k/train.parquet \ 5 | data.val_files=$HOME/data/gsm8k/test.parquet \ 6 | data.train_batch_size=1024 \ 7 | data.val_batch_size=1312 \ 8 | data.max_prompt_length=512 \ 9 | data.max_response_length=512 \ 10 | actor_rollout_ref.model.path=deepseek-ai/deepseek-llm-7b-chat \ 11 | actor_rollout_ref.actor.optim.lr=1e-6 \ 12 | actor_rollout_ref.model.use_remove_padding=True \ 13 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 14 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=32 \ 15 | actor_rollout_ref.actor.ulysses_sequence_parallel_size=2 \ 16 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 17 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 18 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 19 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 20 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=64 \ 21 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 22 | actor_rollout_ref.rollout.name=vllm \ 23 | actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \ 24 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=64 \ 25 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 26 | critic.optim.lr=1e-5 \ 27 | critic.ulysses_sequence_parallel_size=2 \ 28 | critic.model.use_remove_padding=True \ 29 | critic.model.path=deepseek-ai/deepseek-llm-7b-chat \ 30 | critic.model.enable_gradient_checkpointing=True \ 31 | critic.ppo_micro_batch_size_per_gpu=64 \ 32 | critic.model.fsdp_config.param_offload=False \ 33 | critic.model.fsdp_config.grad_offload=False \ 34 | critic.model.fsdp_config.optimizer_offload=False \ 35 | algorithm.kl_ctrl.kl_coef=0.001 \ 36 | trainer.critic_warmup=0 \ 37 | trainer.logger=['console','wandb'] \ 38 | trainer.project_name='verl_example_gsm8k' \ 39 | trainer.experiment_name='deepseek_llm_7b_function_rm_sp2' \ 40 | trainer.n_gpus_per_node=8 \ 41 | trainer.nnodes=1 \ 42 | trainer.save_freq=-1 \ 43 | trainer.test_freq=5 \ 44 | trainer.total_epochs=15 $@ 45 | -------------------------------------------------------------------------------- /tests/verl/utils/dataset/test_rl_dataset.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | import os 15 | import torch 16 | from torch.utils.data import DataLoader 17 | from transformers import AutoTokenizer 18 | 19 | 20 | def get_gsm8k_data(): 21 | # prepare test dataset 22 | url = "https://github.com/eric-haibin-lin/verl-data/raw/refs/heads/main/gsm8k/train.parquet" 23 | local_folder = os.path.expanduser('~/verl-data/gsm8k/') 24 | local_path = os.path.join(local_folder, 'train.parquet') 25 | os.makedirs(local_folder, exist_ok=True) 26 | return local_path 27 | 28 | 29 | def test_rl_dataset(): 30 | from verl.utils.dataset.rl_dataset import RLHFDataset, collate_fn 31 | from verl.utils import hf_tokenizer 32 | tokenizer = hf_tokenizer('deepseek-ai/deepseek-coder-1.3b-instruct') 33 | local_path = get_gsm8k_data() 34 | dataset = RLHFDataset(parquet_files=local_path, tokenizer=tokenizer, prompt_key='prompt', max_prompt_length=256) 35 | 36 | dataloader = DataLoader(dataset=dataset, batch_size=16, shuffle=True, drop_last=True, collate_fn=collate_fn) 37 | 38 | a = next(iter(dataloader)) 39 | 40 | from verl import DataProto 41 | 42 | tensors = {} 43 | non_tensors = {} 44 | 45 | for key, val in a.items(): 46 | if isinstance(val, torch.Tensor): 47 | tensors[key] = val 48 | else: 49 | non_tensors[key] = val 50 | 51 | data_proto = DataProto.from_dict(tensors=tensors, non_tensors=non_tensors) 52 | 53 | data = dataset[0]['input_ids'] 54 | output = tokenizer.batch_decode([data])[0] 55 | print(f'type: type{output}') 56 | print(f'\n\noutput: {output}') 57 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/benchmarks/code_execution.py: -------------------------------------------------------------------------------- 1 | import json 2 | from enum import Enum 3 | from datetime import datetime 4 | from dataclasses import dataclass 5 | 6 | from datasets import load_dataset 7 | 8 | 9 | @dataclass 10 | class CodeExecutionProblem: 11 | question_id: str 12 | contest_id: str 13 | contest_date: datetime 14 | difficulty: str 15 | function_name: str 16 | code: str 17 | input: str 18 | output: str 19 | id: str 20 | problem_id: str 21 | numsteps: int 22 | 23 | def __post_init__(self): 24 | pass 25 | 26 | def insert_output(self, output_list: list[str], pred_list: list[str]) -> dict: 27 | return { 28 | "question_id": self.question_id, 29 | "contest_id": self.contest_id, 30 | "contest_date": self.contest_date.isoformat(), 31 | "difficulty": self.difficulty, 32 | "function_name": self.function_name, 33 | "code": self.code, 34 | "input": self.input, 35 | "output": self.output, 36 | "id": self.id, 37 | "problem_id": self.problem_id, 38 | "numsteps": self.numsteps, 39 | "output_list": output_list, 40 | "pred_list": pred_list, 41 | } 42 | 43 | def insert_output_evaluation( 44 | self, output_list: list[str], code_list: list[str], graded_list: list[bool] 45 | ) -> dict: 46 | output = self.insert_output(output_list, code_list) 47 | output["graded_list"] = graded_list 48 | output["pass@1"] = graded_list.count(True) / len(graded_list) 49 | return output 50 | 51 | def get_evaluation_sample(self) -> dict: 52 | return { 53 | "code": self.code, 54 | "input": self.input, 55 | "output": self.output, 56 | } 57 | 58 | 59 | def load_code_execution_dataset(release_version="release_v1") -> list[CodeExecutionProblem]: 60 | dataset = load_dataset("livecodebench/execution-v2", split="test") 61 | dataset = [CodeExecutionProblem(**p) for p in dataset] # type: ignore 62 | print(f"Loaded {len(dataset)} problems") 63 | return dataset 64 | 65 | 66 | if __name__ == "__main__": 67 | dataset = load_code_execution_dataset() 68 | -------------------------------------------------------------------------------- /verl/single_controller/base/megatron/worker_group.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | 15 | from typing import Dict 16 | 17 | from .worker import DistRankInfo, DistGlobalInfo 18 | from verl.single_controller.base import ResourcePool, WorkerGroup 19 | 20 | 21 | class MegatronWorkerGroup(WorkerGroup): 22 | 23 | def __init__(self, resource_pool: ResourcePool, **kwargs): 24 | super().__init__(resource_pool=resource_pool, **kwargs) 25 | self._megatron_rank_info = None 26 | self._megatron_global_info: DistGlobalInfo = None 27 | 28 | def init_megatron(self, default_megatron_kwargs: Dict = None): 29 | raise NotImplementedError(f"MegatronWorkerGroup.init_megatron should be overwritten") 30 | 31 | def get_megatron_rank_info(self, rank: int) -> DistRankInfo: 32 | assert 0 <= rank < self.world_size, f'rank must be from [0, world_size), Got {rank}' 33 | return self._megatron_rank_info[rank] 34 | 35 | @property 36 | def tp_size(self): 37 | assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized" 38 | return self._megatron_global_info.tp_size 39 | 40 | @property 41 | def dp_size(self): 42 | assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized" 43 | return self._megatron_global_info.dp_size 44 | 45 | @property 46 | def pp_size(self): 47 | assert self._megatron_global_info is not None, "MegatronWorkerGroup._megatron_global_info must be initialized" 48 | return self._megatron_global_info.pp_size 49 | 50 | def get_megatron_global_info(self): 51 | return self._megatron_global_info 52 | -------------------------------------------------------------------------------- /tests/ray/detached_worker/client.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | In client, we can get the server handler and send RPC request 16 | """ 17 | 18 | import ray 19 | import torch 20 | 21 | from verl import DataProto 22 | from verl.single_controller.ray import RayClassWithInitArgs 23 | from verl.single_controller.ray.megatron import NVMegatronRayWorkerGroup 24 | 25 | from tensordict import TensorDict 26 | 27 | from server import Trainer 28 | 29 | 30 | def compute_position_id_with_mask(mask): 31 | return torch.clip(torch.cumsum(mask, dim=-1) - 1, min=0, max=None) 32 | 33 | 34 | if __name__ == '__main__': 35 | 36 | ray.init(address='auto', namespace='verl') 37 | # get the worker group using names 38 | worker_names = ['trainerTrainer_0:0', 'trainerTrainer_0:1'] 39 | cls_with_init_args = RayClassWithInitArgs(cls=Trainer) 40 | worker_group = NVMegatronRayWorkerGroup.from_detached(worker_names=worker_names, 41 | ray_cls_with_init=cls_with_init_args) 42 | 43 | batch_size = 16 44 | sequence_length = 1024 45 | 46 | # give Trainer some data to train 47 | input_ids = torch.randint(low=0, high=256, size=(batch_size, sequence_length), dtype=torch.int64, device='cuda') 48 | attention_mask = torch.ones_like(input_ids) 49 | position_ids = compute_position_id_with_mask(attention_mask) 50 | 51 | data = DataProto(batch=TensorDict( 52 | { 53 | 'input_ids': input_ids, 54 | 'attention_mask': attention_mask, 55 | 'position_ids': position_ids 56 | }, batch_size=batch_size), 57 | meta_info={}) 58 | 59 | output = worker_group.train_model(data) 60 | 61 | print(output) 62 | -------------------------------------------------------------------------------- /verl/utils/megatron/pipeline_parallel.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved. 3 | # 4 | # Licensed under the Apache License, Version 2.0 (the "License"); 5 | # you may not use this file except in compliance with the License. 6 | # You may obtain a copy of the License at 7 | # 8 | # http://www.apache.org/licenses/LICENSE-2.0 9 | # 10 | # Unless required by applicable law or agreed to in writing, software 11 | # distributed under the License is distributed on an "AS IS" BASIS, 12 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 | # See the License for the specific language governing permissions and 14 | # limitations under the License. 15 | 16 | import torch 17 | from megatron.core import parallel_state as mpu 18 | 19 | from .sequence_parallel import pad_to_sequence_parallel 20 | 21 | 22 | def compute_transformers_input_shapes(batches, meta_info): 23 | from flash_attn.bert_padding import unpad_input # flash 2 is a must for Megatron 24 | # pre-compute input shapes for each micro-batch at each pp stage 25 | input_shapes = [] 26 | for model_inputs in batches: 27 | input_ids = model_inputs['input_ids'] 28 | attention_mask = model_inputs['attention_mask'] 29 | input_ids_rmpad = unpad_input(input_ids.unsqueeze(dim=-1), attention_mask)[0] # (total_nnz, 1) 30 | if meta_info['sequence_parallel']: 31 | input_ids_rmpad = pad_to_sequence_parallel(input_ids_rmpad) 32 | # compute shapes for model_inputs 33 | input_shapes.append( 34 | torch.Size([ 35 | input_ids_rmpad.shape[0] // mpu.get_tensor_model_parallel_world_size(), 1, meta_info['hidden_size'] 36 | ])) 37 | else: 38 | # compute shapes for model_inputs 39 | input_shapes.append(torch.Size([input_ids_rmpad.shape[0], 1, meta_info['hidden_size']])) 40 | return input_shapes 41 | 42 | 43 | def make_batch_generator(batches, vpp_size): 44 | if vpp_size > 1: 45 | # has vpp 46 | batch_generator = [batches] * vpp_size # number of vpp chunks 47 | batch_generator = [iter(b) for b in batch_generator] 48 | else: 49 | # no vpp 50 | batch_generator = iter(batches) 51 | return batch_generator 52 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/evaluation/old_results_check.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | import json 3 | from verl.utils.reward_score.livecodebench.lcb_runner.benchmarks import load_generation_dataset, CodeGenerationProblem 4 | from verl.utils.reward_score.livecodebench.lcb_runner.evaluation import codegen_metrics 5 | 6 | 7 | dataset = load_generation_dataset() 8 | 9 | dataset = sorted(dataset, key=lambda x: x.question_id) 10 | 11 | 12 | def check_model(model_key): 13 | path = f"/home/naman/Repos/LiveCodeBench/run_models_outputs/{model_key}/chat_0.2_checked.json" 14 | with open(path) as f: 15 | old_results = json.load(f) 16 | old_results = sorted(old_results, key=lambda x: x["question_id"]) 17 | assert old_results[0]["question_id"] == dataset[0].question_id 18 | 19 | def debug(idx): 20 | codegen_metrics( 21 | [dataset[idx].get_evaluation_sample()], 22 | [old_results[idx]["code_list"][:1]], 23 | debug=True, 24 | ) 25 | 26 | def run(idx): 27 | return codegen_metrics( 28 | [dataset[idx].get_evaluation_sample()], 29 | [old_results[idx]["code_list"]], 30 | ) 31 | 32 | debug(380) 33 | exit() 34 | # debug(196) 35 | # debug(352) 36 | 37 | metrics = codegen_metrics( 38 | [d.get_evaluation_sample() for d in dataset], 39 | [r["code_list"] for r in old_results], 40 | num_process_evaluate=12, 41 | ) 42 | old_pass1 = np.mean([np.mean(r["pass1_list"]) for r in old_results]) 43 | 44 | print(old_pass1) 45 | print(metrics[0]["pass@1"]) 46 | 47 | for idx in range(400): 48 | old_pass1 = np.mean(old_results[idx]["pass1_list"]) 49 | new_pass1 = metrics[0]["detail"]["pass@1"][idx] 50 | if not abs(old_pass1 - new_pass1) < 1e-4: 51 | print(idx, old_pass1, new_pass1) 52 | 53 | 54 | # model_key = "GPT-4-Turbo-1106" 55 | # check_model(model_key) 56 | 57 | model_key = "Claude-3-Opus" 58 | check_model(model_key) 59 | 60 | model_key = "GPT-4-0613" 61 | check_model(model_key) 62 | 63 | model_key = "Mistral-Large" 64 | check_model(model_key) 65 | 66 | model_key = "Claude-3-Sonnet" 67 | check_model(model_key) 68 | 69 | model_key = "GPT-3.5-Turbo-0301" 70 | check_model(model_key) 71 | 72 | model_key = "Gemini-Pro" 73 | check_model(model_key) 74 | -------------------------------------------------------------------------------- /verl/workers/actor/base.py: -------------------------------------------------------------------------------- 1 | # Copyright 2024 Bytedance Ltd. and/or its affiliates 2 | # 3 | # Licensed under the Apache License, Version 2.0 (the "License"); 4 | # you may not use this file except in compliance with the License. 5 | # You may obtain a copy of the License at 6 | # 7 | # http://www.apache.org/licenses/LICENSE-2.0 8 | # 9 | # Unless required by applicable law or agreed to in writing, software 10 | # distributed under the License is distributed on an "AS IS" BASIS, 11 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 | # See the License for the specific language governing permissions and 13 | # limitations under the License. 14 | """ 15 | The base class for Actor 16 | """ 17 | from abc import ABC, abstractmethod 18 | from typing import Iterable, Dict 19 | 20 | from verl import DataProto 21 | import torch 22 | 23 | __all__ = ['BasePPOActor'] 24 | 25 | 26 | class BasePPOActor(ABC): 27 | 28 | def __init__(self, config): 29 | """The base class for PPO actor 30 | 31 | Args: 32 | config (DictConfig): a config passed to the PPOActor. We expect the type to be 33 | DictConfig (https://omegaconf.readthedocs.io/), but it can be any namedtuple in general. 34 | """ 35 | super().__init__() 36 | self.config = config 37 | 38 | @abstractmethod 39 | def compute_log_prob(self, data: DataProto) -> torch.Tensor: 40 | """Compute logits given a batch of data. 41 | 42 | Args: 43 | data (DataProto): a batch of data represented by DataProto. It must contain key ```input_ids```, 44 | ```attention_mask``` and ```position_ids```. 45 | 46 | Returns: 47 | DataProto: a DataProto containing the key ```log_probs``` 48 | 49 | 50 | """ 51 | pass 52 | 53 | @abstractmethod 54 | def update_policy(self, data: DataProto) -> Dict: 55 | """Update the policy with an iterator of DataProto 56 | 57 | Args: 58 | data (DataProto): an iterator over the DataProto that returns by 59 | ```make_minibatch_iterator``` 60 | 61 | Returns: 62 | Dict: a dictionary contains anything. Typically, it contains the statistics during updating the model 63 | such as ```loss```, ```grad_norm```, etc,. 64 | 65 | """ 66 | pass 67 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_qwen2-7b.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | gsm8k_train_path=$HOME/data/gsm8k/train.parquet 4 | gsm8k_test_path=$HOME/data/gsm8k/test.parquet 5 | math_train_path=$HOME/data/math/train.parquet 6 | math_test_path=$HOME/data/math/test.parquet 7 | 8 | train_files="['$gsm8k_train_path', '$math_train_path']" 9 | test_files="['$gsm8k_test_path', '$math_test_path']" 10 | 11 | python3 -m verl.trainer.main_ppo \ 12 | data.train_files="$train_files" \ 13 | data.val_files="$test_files" \ 14 | data.train_batch_size=1024 \ 15 | data.val_batch_size=6312 \ 16 | data.max_prompt_length=1024 \ 17 | data.max_response_length=512 \ 18 | actor_rollout_ref.model.path=Qwen/Qwen2-7B-Instruct \ 19 | actor_rollout_ref.actor.optim.lr=1e-6 \ 20 | actor_rollout_ref.model.use_remove_padding=True \ 21 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 22 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \ 23 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 24 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 25 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 26 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 27 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \ 28 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 29 | actor_rollout_ref.rollout.name=vllm \ 30 | actor_rollout_ref.rollout.gpu_memory_utilization=0.6 \ 31 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \ 32 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 33 | critic.optim.lr=1e-5 \ 34 | critic.model.use_remove_padding=True \ 35 | critic.model.path=Qwen/Qwen2-7B-Instruct \ 36 | critic.model.enable_gradient_checkpointing=True \ 37 | critic.ppo_micro_batch_size_per_gpu=32 \ 38 | critic.model.fsdp_config.param_offload=False \ 39 | critic.model.fsdp_config.grad_offload=False \ 40 | critic.model.fsdp_config.optimizer_offload=False \ 41 | algorithm.kl_ctrl.kl_coef=0.001 \ 42 | trainer.critic_warmup=0 \ 43 | trainer.logger=['console','wandb'] \ 44 | trainer.project_name='verl_example' \ 45 | trainer.experiment_name='Qwen2-7B-Instruct_function_rm' \ 46 | trainer.n_gpus_per_node=8 \ 47 | trainer.nnodes=1 \ 48 | trainer.save_freq=-1 \ 49 | trainer.test_freq=10 \ 50 | trainer.total_epochs=15 $@ 51 | -------------------------------------------------------------------------------- /verl/utils/reward_score/livecodebench/lcb_runner/evaluation/pass_k_utils.py: -------------------------------------------------------------------------------- 1 | import numpy as np 2 | 3 | 4 | def estimate_pass_at_k(num_samples, num_correct, k): 5 | """Estimates pass@k of each problem and returns them in an array.""" 6 | 7 | def estimator(n: int, c: int, k: int) -> float: 8 | """Calculates 1 - comb(n - c, k) / comb(n, k).""" 9 | if n - c < k: 10 | return 1.0 11 | return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1)) 12 | 13 | import itertools 14 | 15 | if isinstance(num_samples, int): 16 | num_samples_it = itertools.repeat(num_samples, len(num_correct)) 17 | else: 18 | assert len(num_samples) == len(num_correct) 19 | num_samples_it = iter(num_samples) 20 | 21 | return np.array( 22 | [estimator(int(n), int(c), k) for n, c in zip(num_samples_it, num_correct)] 23 | ) 24 | 25 | 26 | def compute_metrics_from_results(results, k_list=[1, 5]): 27 | total = [] 28 | correct = [] 29 | task_ids = [] 30 | for task_id, res in results.items(): 31 | all_correct = [] 32 | for generation in res: 33 | gen = np.array(generation) 34 | all_correct.append(np.all(gen > 0)) 35 | task_ids.append(task_id) 36 | total.append(len(all_correct)) 37 | correct.append(sum(all_correct)) 38 | total = np.array(total) 39 | correct = np.array(correct) 40 | ks = k_list 41 | detail_pass_at_k = { 42 | f"pass@{k}": estimate_pass_at_k(total, correct, k).tolist() 43 | for k in ks 44 | if (total >= k).all() 45 | } 46 | pass_at_k = { 47 | f"pass@{k}": estimate_pass_at_k(total, correct, k).mean() 48 | for k in ks 49 | if (total >= k).all() 50 | } 51 | detail_metrics = {k: dict(zip(task_ids, v)) for k, v in detail_pass_at_k.items()} 52 | pass_at_k["detail"] = detail_metrics 53 | return pass_at_k 54 | 55 | 56 | def extract_instance_results(results): 57 | instance_wise_grades = {} 58 | for task_id, res in results.items(): 59 | instance_wise_grades[task_id] = [] 60 | for generation in res: 61 | instance_wise_grades[task_id].append(all([g > 0 for g in generation])) 62 | 63 | instance_wise_grades = [ 64 | v for _, v in sorted(instance_wise_grades.items(), key=lambda item: item[0]) 65 | ] 66 | return instance_wise_grades 67 | -------------------------------------------------------------------------------- /or1_scripts/eval/eval_7b.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | export WORLD_SIZE=${WORLD_SIZE:-1} 5 | export RANK=${RANK:-0} 6 | export MASTER_ADDR=${MASTER_ADDR:-"127.0.0.1"} 7 | export MASTER_PORT=${MASTER_PORT:-29500} 8 | export HYDRA_FULL_ERROR=1 9 | export RAY_BACKEND_LOG_LEVEL=debug 10 | export GPUS_PER_NODE=$(python -c 'import torch; print(torch.cuda.device_count())') 11 | export LIVECODEBENCH_DATA_PATH=${LIVECODEBENCH_DATA_PATH:-./or1_data/eval/livecodebench/livecodebench_2408_2502} 12 | 13 | MODEL_NAME=${MODEL_NAME:-Skywork/Skywork-OR1-Math-7B} 14 | 15 | # Evalation Aime24 16 | python3 -m verl.trainer.main_generation \ 17 | trainer.nnodes=$WORLD_SIZE \ 18 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 19 | model.path=$MODEL_NAME \ 20 | data.path=or1_data/eval/aime24.parquet \ 21 | data.output_path=./outputs/evalation/Aime24_Avg32-Skywork_OR1_Math_7B.pkl \ 22 | data.n_samples=32 \ 23 | data.batch_size=102400 \ 24 | rollout.temperature=0.6 \ 25 | rollout.response_length=32768 \ 26 | rollout.top_k=-1 \ 27 | rollout.top_p=1.0 \ 28 | rollout.gpu_memory_utilization=0.8 \ 29 | rollout.tensor_model_parallel_size=1 30 | 31 | # Evalation Aime25 32 | python3 -m verl.trainer.main_generation \ 33 | trainer.nnodes=$WORLD_SIZE \ 34 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 35 | model.path=$MODEL_NAME \ 36 | data.path=or1_data/eval/aime25.parquet \ 37 | data.output_path=./outputs/evalation/Aime25_Avg32-Skywork_OR1_Math_7B.pkl \ 38 | data.n_samples=32 \ 39 | data.batch_size=102400 \ 40 | rollout.temperature=0.6 \ 41 | rollout.response_length=32768 \ 42 | rollout.top_k=-1 \ 43 | rollout.top_p=1.0 \ 44 | rollout.gpu_memory_utilization=0.8 \ 45 | rollout.tensor_model_parallel_size=1 46 | 47 | # Evalation LiveCodeBench 48 | python3 -m verl.trainer.main_generation \ 49 | trainer.nnodes=$WORLD_SIZE \ 50 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 51 | model.path=$MODEL_NAME \ 52 | data.path=or1_data/eval/livecodebench/livecodebench_2408_2502.parquet \ 53 | data.output_path=./outputs/evalation/LCB_Avg4-Skywork_OR1_Math_7B.pkl \ 54 | data.n_samples=4 \ 55 | data.batch_size=102400 \ 56 | rollout.temperature=0.6 \ 57 | rollout.response_length=32768 \ 58 | rollout.top_k=-1 \ 59 | rollout.top_p=1.0 \ 60 | rollout.gpu_memory_utilization=0.8 \ 61 | rollout.tensor_model_parallel_size=1 -------------------------------------------------------------------------------- /or1_scripts/eval/eval_32b.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | export WORLD_SIZE=${WORLD_SIZE:-1} 5 | export RANK=${RANK:-0} 6 | export MASTER_ADDR=${MASTER_ADDR:-"127.0.0.1"} 7 | export MASTER_PORT=${MASTER_PORT:-29500} 8 | export HYDRA_FULL_ERROR=1 9 | export RAY_BACKEND_LOG_LEVEL=debug 10 | export GPUS_PER_NODE=$(python -c 'import torch; print(torch.cuda.device_count())') 11 | export LIVECODEBENCH_DATA_PATH=${LIVECODEBENCH_DATA_PATH:-./or1_data/eval/livecodebench/livecodebench_2408_2502} 12 | 13 | MODEL_NAME=${MODEL_NAME:-Skywork/Skywork-OR1-32B-Preview} 14 | 15 | # Evalation Aime24 16 | python3 -m verl.trainer.main_generation \ 17 | trainer.nnodes=$WORLD_SIZE \ 18 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 19 | model.path=$MODEL_NAME \ 20 | data.path=or1_data/eval/aime24.parquet \ 21 | data.output_path=./outputs/evalation/Aime24_Avg32-Skywork_OR1_32B_Preview.pkl \ 22 | data.n_samples=32 \ 23 | data.batch_size=102400 \ 24 | rollout.temperature=1.0 \ 25 | rollout.response_length=32768 \ 26 | rollout.top_k=-1 \ 27 | rollout.top_p=1.0 \ 28 | rollout.gpu_memory_utilization=0.8 \ 29 | rollout.tensor_model_parallel_size=2 30 | 31 | # Evalation Aime25 32 | python3 -m verl.trainer.main_generation \ 33 | trainer.nnodes=$WORLD_SIZE \ 34 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 35 | model.path=$MODEL_NAME \ 36 | data.path=or1_data/eval/aime25.parquet \ 37 | data.output_path=./outputs/evalation/Aime25_Avg32-Skywork_OR1_32B_Preview.pkl \ 38 | data.n_samples=32 \ 39 | data.batch_size=102400 \ 40 | rollout.temperature=1.0 \ 41 | rollout.response_length=32768 \ 42 | rollout.top_k=-1 \ 43 | rollout.top_p=1.0 \ 44 | rollout.gpu_memory_utilization=0.8 \ 45 | rollout.tensor_model_parallel_size=2 46 | 47 | # Evalation LiveCodeBench 48 | python3 -m verl.trainer.main_generation \ 49 | trainer.nnodes=$WORLD_SIZE \ 50 | trainer.n_gpus_per_node=$GPUS_PER_NODE \ 51 | model.path=$MODEL_NAME \ 52 | data.path=or1_data/eval/livecodebench/livecodebench_2408_2502.parquet \ 53 | data.output_path=./outputs/evalation/LCB_Avg4-Skywork_OR1_32B_Preview.pkl \ 54 | data.n_samples=4 \ 55 | data.batch_size=102400 \ 56 | rollout.temperature=1.0 \ 57 | rollout.response_length=32768 \ 58 | rollout.top_k=-1 \ 59 | rollout.top_p=1.0 \ 60 | rollout.gpu_memory_utilization=0.8 \ 61 | rollout.tensor_model_parallel_size=2 -------------------------------------------------------------------------------- /tests/e2e/run_qwen_gsm8k_model_rm.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | export VLLM_ATTENTION_BACKEND=XFORMERS 4 | 5 | python3 -m verl.trainer.main_ppo \ 6 | data.train_files=$HOME/data/gsm8k/train.parquet \ 7 | data.val_files=$HOME/data/gsm8k/test.parquet \ 8 | data.train_batch_size=1024 \ 9 | data.val_batch_size=1312 \ 10 | data.max_prompt_length=512 \ 11 | data.max_response_length=512 \ 12 | data.return_raw_chat=True \ 13 | actor_rollout_ref.model.path=Qwen/Qwen2.5-0.5B \ 14 | actor_rollout_ref.actor.optim.lr=1e-6 \ 15 | actor_rollout_ref.model.use_remove_padding=True \ 16 | actor_rollout_ref.actor.optim.lr_warmup_steps_ratio=0.1 \ 17 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 18 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \ 19 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 20 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 21 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 22 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 23 | actor_rollout_ref.rollout.tensor_model_parallel_size=2 \ 24 | actor_rollout_ref.rollout.name=vllm \ 25 | actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \ 26 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 27 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 28 | critic.optim.lr=1e-5 \ 29 | critic.model.use_remove_padding=True \ 30 | critic.optim.lr_warmup_steps_ratio=0.05 \ 31 | critic.model.path=Qwen/Qwen2.5-0.5B \ 32 | critic.model.enable_gradient_checkpointing=False \ 33 | critic.ppo_micro_batch_size_per_gpu=4 \ 34 | critic.model.fsdp_config.param_offload=False \ 35 | critic.model.fsdp_config.grad_offload=False \ 36 | critic.model.fsdp_config.optimizer_offload=False \ 37 | reward_model.enable=True \ 38 | reward_model.model.path=Qwen/Qwen2.5-0.5B\ 39 | reward_model.model.use_remove_padding=True \ 40 | reward_model.model.fsdp_config.param_offload=True \ 41 | reward_model.micro_batch_size_per_gpu=16 \ 42 | algorithm.kl_ctrl.kl_coef=0.001 \ 43 | trainer.critic_warmup=0 \ 44 | trainer.logger=['console'] \ 45 | +trainer.val_before_train=False \ 46 | trainer.project_name='verl_example' \ 47 | trainer.experiment_name='Qwen2.5-0.5B-ci_hybrid_rm' \ 48 | trainer.n_gpus_per_node=8 \ 49 | trainer.nnodes=1 \ 50 | trainer.save_freq=-1 \ 51 | trainer.total_training_steps=1 $@ 52 | -------------------------------------------------------------------------------- /examples/ppo_trainer/run_qwen2.5-32b.sh: -------------------------------------------------------------------------------- 1 | set -x 2 | 3 | gsm8k_train_path=$HOME/data/gsm8k/train.parquet 4 | gsm8k_test_path=$HOME/data/gsm8k/test.parquet 5 | math_train_path=$HOME/data/math/train.parquet 6 | math_test_path=$HOME/data/math/test.parquet 7 | 8 | train_files="['$gsm8k_train_path', '$math_train_path']" 9 | test_files="['$gsm8k_test_path', '$math_test_path']" 10 | 11 | python3 -m verl.trainer.main_ppo \ 12 | data.train_files="$train_files" \ 13 | data.val_files="$test_files" \ 14 | data.train_batch_size=1024 \ 15 | data.val_batch_size=6304 \ 16 | data.max_prompt_length=1024 \ 17 | data.max_response_length=1024 \ 18 | actor_rollout_ref.model.path=Qwen/Qwen2.5-32B-Instruct \ 19 | actor_rollout_ref.model.enable_gradient_checkpointing=False \ 20 | actor_rollout_ref.actor.optim.lr=1e-6 \ 21 | actor_rollout_ref.model.use_remove_padding=True \ 22 | actor_rollout_ref.actor.ppo_mini_batch_size=256 \ 23 | actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 \ 24 | actor_rollout_ref.model.enable_gradient_checkpointing=True \ 25 | actor_rollout_ref.actor.fsdp_config.param_offload=False \ 26 | actor_rollout_ref.actor.fsdp_config.grad_offload=False \ 27 | actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \ 28 | actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \ 29 | actor_rollout_ref.rollout.tensor_model_parallel_size=4 \ 30 | actor_rollout_ref.rollout.name=vllm \ 31 | actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \ 32 | actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \ 33 | actor_rollout_ref.ref.fsdp_config.param_offload=True \ 34 | critic.optim.lr=1e-5 \ 35 | critic.model.use_remove_padding=True \ 36 | critic.model.path=Qwen/Qwen2.5-32B-Instruct \ 37 | critic.model.enable_gradient_checkpointing=False \ 38 | critic.ppo_micro_batch_size_per_gpu=8 \ 39 | critic.model.fsdp_config.param_offload=False \ 40 | critic.model.fsdp_config.grad_offload=False \ 41 | critic.model.fsdp_config.optimizer_offload=False \ 42 | algorithm.kl_ctrl.kl_coef=0.0001 \ 43 | trainer.critic_warmup=0 \ 44 | trainer.logger=['console','wandb'] \ 45 | trainer.project_name='verl_example' \ 46 | trainer.experiment_name='Qwen2.5-32B-Instruct_function_rm' \ 47 | trainer.n_gpus_per_node=8 \ 48 | trainer.nnodes=4 \ 49 | trainer.save_freq=-1 \ 50 | trainer.test_freq=10 \ 51 | trainer.total_epochs=15 $@ 52 | --------------------------------------------------------------------------------