[misc] refactor: refactor the folders structure

.github/workflows/sanity.yml

@@ -35,6 +35,7 @@ jobs:
       - name: Install the current repository
         run: |
           pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu
+          pip3 install --no-deps "verl @ git+https://github.com/verl-project/verl.git@main"
           pip3 install -r requirements.txt
           pip3 install -r requirements-test.txt
           pip3 install --no-deps -e .

.github/workflows/type-coverage-check.yml

@@ -21,11 +21,12 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cpu
+          pip3 install --no-deps "verl @ git+https://github.com/verl-project/verl.git@main"
           pip3 install -r requirements.txt
           pip3 install --no-deps -e .
       - name: Run type annotation coverage check
         run: |
           python3 tests/special_sanity/type_coverage_check.py
       - name: Run docstring coverage check
         run: |
-          python3 tests/special_sanity/check_api_docs.py verl
+          python3 tests/special_sanity/check_api_docs.py verl_omni

.pre-commit-config.yaml

@@ -15,7 +15,7 @@ repos:
   - repo: local
     hooks:
       - id: autogen-trainer-cfg
-        name: Generate and verify verl/trainer/config/_generated_*.yaml
+        name: Generate and verify verl_omni/trainer/config/_generated_*.yaml
         entry: scripts/generate_trainer_config.sh
         language: script
         pass_filenames: false

examples/flowgrpo_trainer/data_process/qwenimage_ocr.py

@@ -20,7 +20,6 @@
 import os
 
 import datasets
-
 from verl.utils.hdfs_io import copy, makedirs
 
 

examples/flowgrpo_trainer/diffusers_impl/qwen_image.py

@@ -22,11 +22,11 @@
 from diffusers.models.transformers.transformer_qwenimage import QwenImageTransformer2DModel
 from diffusers.pipelines.qwenimage.pipeline_qwenimage import calculate_shift
 from tensordict import TensorDict
-
-from verl.models.diffusion_model import DiffusionModelBase
 from verl.utils import tensordict_utils as tu
 from verl.utils.device import get_device_name
-from verl.workers.config import DiffusionModelConfig
+
+from verl_omni.models.diffusion_model import DiffusionModelBase
+from verl_omni.workers.config import DiffusionModelConfig
 
 from ..scheduler import FlowMatchSDEDiscreteScheduler
 

examples/flowgrpo_trainer/reward_fn.py

@@ -72,7 +72,6 @@ async def compute_score_ocr(
     import re
 
     import Levenshtein
-
     from verl.utils.experimental.reward_utils import pil_image_to_base64
     from verl.utils.ray_utils import get_event_loop
 

examples/flowgrpo_trainer/run_qwen_image_ocr_lora.sh

@@ -11,7 +11,7 @@ reward_path=examples/flowgrpo_trainer/reward_fn.py
 reward_model_name=$HOME/models/Qwen/Qwen3-VL-8B-Instruct
 
 
-python3 -m verl.trainer.main_flowgrpo \
+python3 -m verl_omni.trainer.main_flowgrpo \
     algorithm.adv_estimator=flow_grpo \
     data.train_files=$ocr_train_path \
     data.val_files=$ocr_test_path \

examples/flowgrpo_trainer/vllm_omni_impl/pipeline_qwenimage.py

@@ -20,7 +20,7 @@
 from vllm_omni.diffusion.models.qwen_image import QwenImagePipeline
 from vllm_omni.diffusion.request import OmniDiffusionRequest
 
-from verl.models.diffusion_model import VllmOmniPipelineBase
+from verl_omni.models.diffusion_model import VllmOmniPipelineBase
 
 from ..scheduler import FlowMatchSDEDiscreteScheduler
 

pyproject.toml

@@ -13,7 +13,7 @@ build-backend = "setuptools.build_meta"
 # -------------------------------
 [project]
 name = "verl-omni"
-# We'll mark the version as "dynamic" because it's read from the file "verl/version/version" 
+# We'll mark the version as "dynamic" because it's read from the file "verl_omni/version/version" 
 # (PEP 621 calls this "dynamic version"). 
 # The actual version is specified in the [tool.setuptools.dynamic] section below.
 dynamic = ["version", "dependencies", "optional-dependencies", "authors", "urls"]
@@ -33,7 +33,7 @@ line-length = 120
 exclude = ["scripts/legacy_model_merger.py"]
 
 [tool.ruff.lint]
-isort = {known-first-party = ["verl"]}
+isort = {known-first-party = ["verl_omni"]}
 # c.f. https://github.com/vllm-project/vllm/blob/ce8d6b75fc0586045df75ee1568a5b5f9957251b/pyproject.toml
 select = [
     # pycodestyle
@@ -79,11 +79,7 @@ ignore_errors = true
 
 [[tool.mypy.overrides]]
 module = [
-"verl.trainer.config.algorithm",
-"verl.trainer.ppo.core_algos",
-"verl.trainer.ppo.reward",
-"verl.workers.reward_manager",
-"verl.workers.reward_manager.*",
+"verl_omni.trainer.config.algorithm",
 ]
 ignore_errors = false
 
@@ -95,20 +91,19 @@ ignore_errors = false
 # This corresponds to `include_package_data=True` in setup.py.
 include-package-data = true
 
-# We read the version from a file in 'verl/version/version'
+# We read the version from a file in 'verl_omni/version/version'
 [tool.setuptools.dynamic]
-version = {file = "verl/version/version"}
+version = {file = "verl_omni/version/version"}
 
 # If you need to mimic `package_dir={'': '.'}`:
 [tool.setuptools.package-dir]
 "" = "."
 
 # If you need to include specific non-Python data (like YAML files or version file):
-# This is the rough equivalent of package_data={'': ['version/*'], 'verl': ['trainer/config/*.yaml']}
 [tool.setuptools.package-data]
-verl = [
+verl_omni = [
   "version/*",
   "trainer/config/*.yaml",
   "trainer/config/*/*.yaml",
-  "experimental/*/config/*.yaml",
+  "trainer/config/*/*/*.yaml",
 ]

requirements.txt

@@ -1,7 +1,10 @@
 # requirements.txt records the full set of dependencies for development
+verl>=0.7.1
 accelerate
+cachetools
 codetiming
 datasets
+diffusers>=0.37.1
 dill
 hydra-core
 liger-kernel
@@ -16,7 +19,6 @@ ray[default]
 tensordict>=0.8.0,<=0.10.0,!=0.9.0
 torchdata
 transformers
-# vllm==0.8.4
 wandb
 packaging>=20.0
 uvicorn

scripts/generate_trainer_config.sh

@@ -4,10 +4,6 @@ set -euox pipefail
 
 # Define config specifications: "config_name:output_file:config_arg"
 CONFIG_SPECS=(
-    "ppo_trainer:_generated_ppo_trainer.yaml:"
-    "ppo_trainer:_generated_ppo_megatron_trainer.yaml:model_engine=megatron"
-    "ppo_trainer:_generated_ppo_veomni_trainer.yaml:model_engine=veomni"
-    "ppo_trainer:_generated_ppo_torchtitan_trainer.yaml:model_engine=torchtitan"
     "diffusion_trainer:_generated_diffusion_trainer.yaml:--config-name=diffusion_trainer.yaml"
 )
 
@@ -16,12 +12,12 @@ generate_config() {
     local output_file="$2"
     local config_arg="$3"
 
-    local target_cfg="verl/trainer/config/${output_file}"
+    local target_cfg="verl_omni/trainer/config/${output_file}"
     local tmp_header=$(mktemp)
     local tmp_cfg=$(mktemp)
 
     echo "# This reference configration yaml is automatically generated via 'scripts/generate_trainer_config.sh'" > "$tmp_header"
-    echo "# in which it invokes 'python3 scripts/print_cfg.py --cfg job ${config_arg}' to flatten the 'verl/trainer/config/${config_name}.yaml' config fields into a single file." >> "$tmp_header"
+    echo "# in which it invokes 'python3 scripts/print_cfg.py --cfg job ${config_arg}' to flatten the 'verl_omni/trainer/config/${config_name}.yaml' config fields into a single file." >> "$tmp_header"
     echo "# Do not modify this file directly." >> "$tmp_header"
     echo "# The file is usually only for reference and never used." >> "$tmp_header"
     echo "" >> "$tmp_header"
@@ -43,7 +39,7 @@ done
 
 for spec in "${CONFIG_SPECS[@]}"; do
     IFS=':' read -r config_name output_file config_arg <<< "$spec"
-    target_cfg="verl/trainer/config/${output_file}"
+    target_cfg="verl_omni/trainer/config/${output_file}"
     if ! git diff --exit-code -- "$target_cfg" >/dev/null; then
         echo "✖ $target_cfg is out of date. Please regenerate via 'scripts/generate_trainer_config.sh' and commit the changes."
         exit 1

scripts/print_cfg.py

@@ -17,18 +17,14 @@
     raise ImportError("Please install hydra-core via 'pip install hydra-core' and retry.") from e
 
 
-@hydra.main(config_path="../verl/trainer/config", config_name="ppo_trainer", version_base=None)
+@hydra.main(config_path="../verl_omni/trainer/config", config_name="diffusion_trainer", version_base=None)
 def main(config):
-    """Main entry point for PPO training with Hydra configuration management.
+    """Main entry point for printing the resolved diffusion trainer config.
 
     Args:
         config_dict: Hydra configuration dictionary containing training parameters.
     """
     print(config)
-    from verl.utils.config import omega_conf_to_dataclass
-
-    profiler_config = omega_conf_to_dataclass(config.critic.profiler)
-    print(profiler_config)
 
 
 if __name__ == "__main__":

setup.py

@@ -20,13 +20,15 @@
 
 version_folder = os.path.dirname(os.path.join(os.path.abspath(__file__)))
 
-with open(os.path.join(version_folder, "verl/version/version")) as f:
+with open(os.path.join(version_folder, "verl_omni/version/version")) as f:
     __version__ = f.read().strip()
 
 install_requires = [
     "accelerate",
+    "cachetools",
     "codetiming",
     "datasets",
+    "diffusers",
     "dill",
     "hydra-core",
     "numpy<2.0.0",
@@ -45,57 +47,37 @@
 ]
 
 TEST_REQUIRES = ["pytest", "pre-commit", "py-spy", "pytest-asyncio", "pytest-rerunfailures"]
-PRIME_REQUIRES = ["pyext"]
-GEO_REQUIRES = ["mathruler", "torchvision", "qwen_vl_utils"]
-GPU_REQUIRES = ["liger-kernel", "flash-attn"]
-MATH_REQUIRES = ["math-verify"]  # Add math-verify as an optional dependency
+GPU_REQUIRES = ["flash-attn"]
 VLLM_REQUIRES = ["tensordict>=0.8.0,<=0.10.0,!=0.9.0", "vllm>=0.8.5,<=0.12.0"]
-TRTLLM_REQUIRES = ["tensorrt-llm>=1.2.0rc6"]
-SGLANG_REQUIRES = [
-    "tensordict>=0.8.0,<=0.10.0,!=0.9.0",
-    "sglang[srt,openai]==0.5.8",
-    "torch==2.9.1",
-]
-TRL_REQUIRES = ["trl<=0.9.6"]
-MCORE_REQUIRES = ["mbridge"]
-TRANSFERQUEUE_REQUIRES = ["TransferQueue==0.1.6"]
 
 extras_require = {
     "test": TEST_REQUIRES,
-    "prime": PRIME_REQUIRES,
-    "geo": GEO_REQUIRES,
     "gpu": GPU_REQUIRES,
-    "math": MATH_REQUIRES,
     "vllm": VLLM_REQUIRES,
-    "sglang": SGLANG_REQUIRES,
-    "trl": TRL_REQUIRES,
-    "mcore": MCORE_REQUIRES,
-    "trtllm": TRTLLM_REQUIRES,
-    "transferqueue": TRANSFERQUEUE_REQUIRES,
 }
 
 
 this_directory = Path(__file__).parent
 long_description = (this_directory / "README.md").read_text()
 
 setup(
-    name="verl",
+    name="verl-omni",
     version=__version__,
     package_dir={"": "."},
-    packages=find_packages(where="."),
-    url="https://github.com/verl-project/verl",
+    packages=find_packages(where=".", include=["verl_omni", "verl_omni.*"]),
+    url="https://github.com/verl-project/verl-omni",
     license="Apache 2.0",
     author="Bytedance - Seed - MLSys",
-    author_email="zhangchi.usc1992@bytedance.com, gmsheng@connect.hku.hk",
-    description="verl: Volcano Engine Reinforcement Learning for LLM",
+    author_email="yhuangch@cse.ust.hk",
+    description="verl-omni: Easy, fast, and stable RL training for diffusion and omni-modality models",
     install_requires=install_requires,
     extras_require=extras_require,
     package_data={
         "": ["version/*"],
-        "verl": [
+        "verl_omni": [
             "trainer/config/*.yaml",
             "trainer/config/*/*.yaml",
-            "experimental/*/config/*.yaml",
+            "trainer/config/*/*/*.yaml",
         ],
     },
     include_package_data=True,

tests/experimental/agent_loop/conftest.py

@@ -13,6 +13,13 @@
 # limitations under the License.
 import pytest
 
+# TODO (mike): to be dropped once `verl` drops its legacy diffusion
+# implementations.
+try:
+    import verl_omni  # noqa: F401
+except ImportError:
+    pass
+
 
 def pytest_configure(config):
     config.addinivalue_line("markers", "vllm_omni: requires the vllm-omni package")

tests/experimental/agent_loop/test_diffusion_agent_loop.py

@@ -19,7 +19,6 @@
 import pytest
 import ray
 from omegaconf import DictConfig
-
 from verl.experimental.agent_loop.agent_loop import AgentLoopManager
 from verl.protocol import DataProto
 
@@ -59,7 +58,7 @@ def _create_tp_compatible_model(parent_dir, src_model_path, num_attention_heads=
 def init_config() -> DictConfig:
     from hydra import compose, initialize_config_dir
 
-    with initialize_config_dir(config_dir=os.path.abspath("verl/trainer/config")):
+    with initialize_config_dir(config_dir=os.path.abspath("verl_omni/trainer/config")):
         config = compose(config_name="diffusion_trainer")
 
     base_model_path = os.path.expanduser("~/models/tiny-random/Qwen-Image")

tests/experimental/reward_loop/test_visual_reward_manager.py

@@ -17,7 +17,6 @@
 import ray
 import torch
 from hydra import compose, initialize_config_dir
-
 from verl.experimental.reward_loop import RewardLoopManager
 from verl.protocol import DataProto
 from verl.utils import hf_tokenizer
@@ -65,8 +64,8 @@ def test_reward_model_genrm():
             }
         }
     )
-    with initialize_config_dir(config_dir=os.path.abspath("verl/trainer/config")):
-        config = compose(config_name="ppo_trainer")
+    with initialize_config_dir(config_dir=os.path.abspath("verl_omni/trainer/config")):
+        config = compose(config_name="diffusion_trainer")
 
     rollout_model_name = os.path.expanduser("~/models/tiny-random/Qwen-Image")
     reward_model_name = os.path.expanduser("~/models/tiny-random/qwen3-vl")
@@ -118,8 +117,8 @@ def test_rule_reward():
             }
         }
     )
-    with initialize_config_dir(config_dir=os.path.abspath("verl/trainer/config")):
-        config = compose(config_name="ppo_trainer")
+    with initialize_config_dir(config_dir=os.path.abspath("verl_omni/trainer/config")):
+        config = compose(config_name="diffusion_trainer")
 
     rollout_model_name = os.path.expanduser("~/models/tiny-random/Qwen-Image")
 

tests/models/test_diffusers_fsdp_engine.py

@@ -19,16 +19,17 @@
 import pytest
 import ray
 import torch
-
 from verl import DataProto
-from verl.models.diffusion_model import build_scheduler
 from verl.single_controller.ray import RayClassWithInitArgs, RayResourcePool, RayWorkerGroup
 from verl.utils import tensordict_utils as tu
-from verl.workers.config import DiffusionModelConfig, FSDPDiffusionActorConfig, TrainingWorkerConfig
-from verl.workers.engine_workers import TrainingWorker
-from verl.workers.utils.losses import diffusion_loss
+from verl.workers.config import TrainingWorkerConfig
 from verl.workers.utils.padding import embeds_padding_2_no_padding
 
+from verl_omni.models.diffusion_model import build_scheduler
+from verl_omni.workers.config import DiffusionModelConfig, FSDPDiffusionActorConfig
+from verl_omni.workers.engine_workers import TrainingWorker
+from verl_omni.workers.utils.losses import diffusion_loss
+
 EXTERNAL_LIB = "examples.flowgrpo_trainer.diffusers_impl.qwen_image"
 
 
@@ -44,10 +45,9 @@ def create_training_config(model_type, strategy, device_count, model):
 
     if strategy in ["fsdp", "fsdp2"]:
         from hydra import compose, initialize_config_dir
-
         from verl.utils.config import omega_conf_to_dataclass
 
-        with initialize_config_dir(config_dir=os.path.abspath("verl/trainer/config/diffusion/model")):
+        with initialize_config_dir(config_dir=os.path.abspath("verl_omni/trainer/config/diffusion/model")):
             cfg = compose(
                 config_name="diffusion_model",
                 overrides=[
@@ -63,7 +63,7 @@ def create_training_config(model_type, strategy, device_count, model):
             )
         model_config: DiffusionModelConfig = omega_conf_to_dataclass(cfg)
 
-        with initialize_config_dir(config_dir=os.path.abspath("verl/trainer/config/diffusion/actor")):
+        with initialize_config_dir(config_dir=os.path.abspath("verl_omni/trainer/config/diffusion/actor")):
             cfg = compose(
                 config_name="dp_diffusion_actor",
                 overrides=[