[tests] BREAKING: move recipe.dapo.src to recipe.dapo; move test files to their own namespaces (tests/verl/xxx -> tests/xxx)

.github/workflows/dataset.yml

@@ -49,10 +49,10 @@ jobs:
         run: |
           [ ! -d "$HOME/verl-data" ] && git clone --depth 1 https://github.com/eric-haibin-lin/verl-data ~/verl-data
           python3 examples/data_preprocess/geo3k.py
-          pytest -s -x tests/verl/utils/dataset/test_rl_dataset.py
-          pytest -s -x tests/verl/utils/dataset/test_sft_dataset.py
-          pytest -s -x tests/verl/utils/test_import_utils.py
-      #          pytest -s -x tests/verl/utils/dataset/test_rm_dataset.py
+          pytest -s -x tests/utils/dataset/test_rl_dataset.py
+          pytest -s -x tests/utils/dataset/test_sft_dataset.py
+          pytest -s -x tests/utils/test_import_utils.py
+          # pytest -s -x tests/utils/dataset/test_rm_dataset.py
       - name: Running ray test using cupy (move it to L20 when dockerfile ready)
         run: |
           cd tests/ray_gpu

.github/workflows/model.yml

@@ -15,8 +15,8 @@ on:
       - "verl/**/*.py"
       # Entrypoints
       - ".github/workflows/model.yml"
-      - "tests/checkpoint/test_fsdp_ckpt.py"
-      - "tests/model/test_transformers_ulysses.py"
+      - "tests/utils/checkpoint/test_fsdp_ckpt.py"
+      - "tests/models/test_transformers_ulysses.py"
       - "tests/distributed/run_all.sh"
 
 # Declare permissions just read content.
@@ -46,37 +46,37 @@ jobs:
           pip3 install --upgrade transformers
       - name: Running rmpad model tests on 8 L20 GPUs + flash_attn 2.5.8
         run: |
-          pytest -s tests/model/test_transformer.py
+          pytest -s tests/models/test_transformer.py
       - name: Running rmpad model tests on 8 L20 GPUs + latest flash_attn
         run: |
           pip3 install --upgrade flash_attn --no-build-isolation
-          pytest -s tests/model/test_transformer.py
+          pytest -s tests/models/test_transformer.py
       - name: Running FSDP rmpad model tests on 8 L20 GPUs + latest flash_attn
         run: |
-          STRATEGY=fsdp torchrun --nproc_per_node=8 tests/checkpoint/test_fsdp_ckpt.py
+          STRATEGY=fsdp torchrun --nproc_per_node=8 tests/utils/checkpoint/test_fsdp_ckpt.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + latest transformers
         run: |
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + transformers 4.49.0
         run: |
           pip3 install transformers==4.49.0
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + transformers 4.48.0
         run: |
           pip3 install transformers==4.48.0
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + transformers 4.47.0
         run: |
           pip3 install transformers==4.47.0
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + transformers 4.46.0
         run: |
           pip3 install transformers==4.46.0
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Running transformers ulysses tests on 8 L20 GPUs + transformers 4.45.0
         run: |
           pip3 install transformers==4.45.0
-          torchrun --nproc_per_node=8 -m pytest tests/model/test_transformers_ulysses.py
+          torchrun --nproc_per_node=8 -m pytest tests/models/test_transformers_ulysses.py
       - name: Run distributed test
         run: |
           bash tests/distributed/run_all.sh
@@ -106,4 +106,4 @@ jobs:
       - name: Running FSDP2 rmpad model tests on 8 L20 GPUs + latest flash_attn
         run: |
           pip3 install --upgrade flash_attn --no-build-isolation
-          STRATEGY=fsdp2 torchrun --nproc_per_node=8 tests/checkpoint/test_fsdp_ckpt.py
+          STRATEGY=fsdp2 torchrun --nproc_per_node=8 tests/utils/checkpoint/test_fsdp_ckpt.py

.github/workflows/sgl.yml

@@ -68,13 +68,13 @@ jobs:
           pip3 install -e .[test,gpu,sglang] --no-deps
       - name: Test the latest SGLang
         run: |
-          cd tests/rollout
+          cd tests/workers/rollout
           torchrun --nnodes=1 --nproc_per_node=4 $(which pytest) -s test_sglang_spmd.py
       - name: Test the latest SGLang async
         run: |
-          cd tests/rollout
+          cd tests/workers/rollout
           torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_async_spmd.py
       - name: Test the latest SGLang Rollout async with tool
         run: |
-          cd tests/rollout
+          cd tests/workers/rollout
           torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_async_rollout_w_tools.py

.github/workflows/verl_unit_test.yml

@@ -43,9 +43,9 @@ jobs:
           pip install -e .[test]
       - name: Running test protocol.py
         run: |
-          cd tests/verl
+          cd tests
           pytest -s -x test_protocol.py
-      - name: Running utils tests
+      - name: running utils tests
         run: |
-          cd tests/verl/utils
-          pytest -s -x --ignore=dataset/ --ignore=test_torch_functional.py .
+          cd tests/utils
+          pytest -s -x --ignore=dataset/ --ignore=checkpoint/ --ignore=test_flops_counter.py --ignore=test_torch_functional.py .

.github/workflows/vllm.yml

@@ -28,7 +28,7 @@ on:
       - "!**/*sglang*"
       # Entrypoints
       - ".github/workflows/vllm.yml"
-      - "tests/generation"
+      - "tests/e2e/generation"
       - "tests/rollout"
       - "verl/trainer/main_generation.py"
       - "verl/trainer/config/generation.yaml"
@@ -72,26 +72,26 @@ jobs:
         # Disable requests to avoid network errors
       - name: Running vllm tests on 8 L20 GPUs
         run: |
-          cd tests/rollout
+          cd tests/workers/rollout
           torchrun --standalone --nnodes=1 --nproc_per_node=8 $(which pytest) -s test_vllm_hf_loader.py
       - name: Test the latest vLLM
         run: |
           pip3 install --upgrade vllm==0.7.3
-          cd tests/rollout
+          cd tests/workers/rollout
           torchrun --standalone --nnodes=1 --nproc_per_node=4 $(which pytest) -s test_vllm_spmd.py
       - name: Run Qwen 0.5B generation test
         run: |
-          cd tests/generation
+          cd tests/e2e/generation
           export OUTPUT_PATH="${HOME}/data/gen/qwen_05_gen_test.parquet"
           MODEL_ID=Qwen/Qwen2.5-0.5B-Instruct NGPUS_PER_NODE=4 GEN_TP=2 bash ./run_gen_qwen05.sh
           rm -rf "${OUTPUT_PATH}"
       - name: Run Qwen 0.5B generation test when world_size == 1
         run: |
-          cd tests/generation
+          cd tests/e2e/generation
           export OUTPUT_PATH="${HOME}/data/gen/qwen_05_gen_test.parquet"
           MODEL_ID=Qwen/Qwen2.5-0.5B-Instruct NGPUS_PER_NODE=1 GEN_TP=1 bash ./run_gen_qwen05.sh
           rm -rf "${OUTPUT_PATH}"
       - name: Running multi-turn rollout tests on 8 L20 GPUs
         run: |
           pip3 install --upgrade vllm==0.8.3 tensordict==0.7.2
-          python3 tests/rollout/test_vllm_multi_turn.py
+          python3 tests/workers/rollout/test_vllm_multi_turn.py

.gitignore

@@ -109,9 +109,6 @@ ENV/
 # Mac
 .DS_Store
 
-# output logs
-tests/e2e/toy_examples/deepspeed/synchronous/output.txt
-
 # vim
 *.swp
 

docs/api/trainer.rst

@@ -0,0 +1,21 @@
+Trainers
+=========================
+
+Trainers drive the training loop. Introducing new trainer classes in case of new training paradiam is encouraged.
+
+.. autosummary::
+   :nosignatures:
+
+   verl.trainer.ppo.ray_trainer.RayPPOTrainer
+
+
+Core APIs
+~~~~~~~~~~~~~~~~~
+
+.. autoclass::  verl.trainer.ppo.ray_trainer.RayPPOTrainer
+
+.. automodule:: verl.utils.tokenizer
+   :members: hf_tokenizer
+
+.. automodule:: verl.single_controller
+   :members: Worker, WorkerGroup, ClassWithInitArgs, ResourcePool

docs/perf/perf_tuning.rst

@@ -51,12 +51,12 @@ Currently, for llama, mistral, gemma1 and qwen based models, users can enable `u
 sequence packing implementation provided by transformers library.
 
 For other models, transformers library may also support it but we haven't tested it yet.
-Users can add the desired model config to the  `test_transformer.py <https://github.com/volcengine/verl/blob/main/tests/model/test_transformer.py#L24>`_ file.
+Users can add the desired model config to the  `test_transformer.py <https://github.com/volcengine/verl/blob/main/tests/models/test_transformer.py#L24>`_ file.
 And test its functionaility by running the following command:
 
 .. code-block:: bash
 
-  pytest -s tests/model/test_transformer.py
+  pytest -s tests/models/test_transformer.py
 
 If the test passes, you can add your desired model into the model `registry.py <https://github.com/volcengine/verl/blob/main/verl/models/registry.py#L24>`_ file.
 Then, you can enjoy the performance boost of sequence packing

recipe/dapo/run_dapo_early_qwen2.5_32b.sh

@@ -58,7 +58,7 @@ gen_tp=4
 
 ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     --working-dir "${WORKING_DIR}" \
-    -- python3 -m recipe.dapo.src.main_dapo \
+    -- python3 -m recipe.dapo.main_dapo \
     data.train_files="${TRAIN_FILE}" \
     data.val_files="${TEST_FILE}" \
     data.prompt_key=prompt \
@@ -125,4 +125,4 @@ ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     trainer.save_freq=5 \
     trainer.total_epochs=1 \
     trainer.default_local_dir="${CKPTS_DIR}" \
-    trainer.resume_mode=auto
+    trainer.resume_mode=auto

recipe/dapo/run_dapo_qwen2.5_32b.sh

@@ -58,7 +58,7 @@ gen_tp=4
 
 ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     --working-dir "${WORKING_DIR}" \
-    -- python3 -m recipe.dapo.src.main_dapo \
+    -- python3 -m recipe.dapo.main_dapo \
     data.train_files="${TRAIN_FILE}" \
     data.val_files="${TEST_FILE}" \
     data.prompt_key=prompt \
@@ -127,4 +127,4 @@ ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     trainer.save_freq=5 \
     trainer.total_epochs=1 \
     trainer.default_local_dir="${CKPTS_DIR}" \
-    trainer.resume_mode=auto
+    trainer.resume_mode=auto

recipe/dapo/run_dapo_wo_ds_qwen2.5_32b.sh

@@ -56,7 +56,7 @@ gen_tp=4
 
 ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     --working-dir "${WORKING_DIR}" \
-    -- python3 -m recipe.dapo.src.main_dapo \
+    -- python3 -m recipe.dapo.main_dapo \
     data.train_files="${TRAIN_FILE}" \
     data.val_files="${TEST_FILE}" \
     data.prompt_key=prompt \
@@ -122,4 +122,4 @@ ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     trainer.save_freq=5 \
     trainer.total_epochs=1 \
     trainer.default_local_dir="${CKPTS_DIR}" \
-    trainer.resume_mode=auto
+    trainer.resume_mode=auto

recipe/dapo/test_dapo_7b.sh

@@ -55,7 +55,7 @@ offload=False
 
 ray job submit --no-wait --runtime-env="${RUNTIME_ENV}" \
     --working-dir "${WORKING_DIR}" \
-    -- python3 -m recipe.dapo.src.main_dapo \
+    -- python3 -m recipe.dapo.main_dapo \
     data.train_files="${TRAIN_FILE}" \
     data.val_files="${TEST_FILE}" \
     data.prompt_key=prompt \

tests/e2e/run_dapo.sh

@@ -41,7 +41,7 @@ gen_prompt_bsz=$((train_prompt_bsz * 4))
 
 exp_name="$(basename "${MODEL_ID,,}")-dapo-minimal"
 
-python3 -m recipe.dapo.src.main_dapo \
+python3 -m recipe.dapo.main_dapo \
     data.train_files="${HOME}/data/gsm8k/train.parquet" \
     data.val_files="${HOME}/data/gsm8k/test.parquet" \
     reward_model.reward_manager=dapo \

tests/utility/test_timeout_decorator.py

@@ -12,17 +12,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import time
-import os
-import sys
 import multiprocessing
-import queue
-import pytest # Import pytest
-from functools import wraps
+import sys
+import threading
+import time
 
+import pytest  # Import pytest
 
 from verl.utils.py_functional import timeout_limit as timeout
-import threading
 
 # --- Test Task Functions ---
 TEST_TIMEOUT_SECONDS = 1.5 # Timeout duration for tests

tests/verl/test_flops_counter.py → tests/utils/test_flops_counter.py

@@ -12,9 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import pytest
 import math
-import json
+
+import pytest
+
 from verl.utils.flops_counter import FlopsCounter
 
 VALID_CONFIG_TYPE = {"llama", "qwen2", "qwen3", "qwen3_moe", "deepseek_v3"}

tests/rollout/test_vllm_multi_turn.py → tests/workers/rollout/test_vllm_multi_turn.py

@@ -21,7 +21,7 @@
 from openai.types.chat.chat_completion import ChatCompletion
 from vllm.entrypoints.openai.protocol import ChatCompletionRequest, ChatCompletionResponse, ChatCompletionStreamResponse, ErrorResponse
 
-from tests.rollout.async_rollout_utils import init_async_rollout_manager
+from tests.workers.rollout.async_rollout_utils import init_async_rollout_manager
 from verl.protocol import DataProto
 
 

verl/single_controller/base/worker.py

@@ -136,6 +136,7 @@ def _configure_before_init(self, register_center_name: str, rank: int):
     def __init__(self, cuda_visible_devices=None) -> None:
         # construct a meta from environment variable. Note that the import must be inside the class because it is executed remotely
         import os
+
         import torch
         from packaging import version
 

verl/utils/megatron_utils.py

@@ -31,9 +31,9 @@
 from megatron.core.utils import get_attr_wrapped_model
 from transformers import PretrainedConfig
 
-from verl.utils.torch_dtypes import PrecisionType
-from verl.utils.model import normalize_model_name
 import verl.utils.megatron.tensor_parallel as tp_utils
+from verl.utils.model import normalize_model_name
+from verl.utils.torch_dtypes import PrecisionType
 
 
 def get_model_config(model):

verl/utils/reward_score/math_dapo.py

@@ -14,9 +14,9 @@
 # Adapted from https://github.com/EleutherAI/lm-evaluation-harness/blob/main/lm_eval/tasks/hendrycks_math/utils.py
 
 import re
-import signal
 from typing import Optional
 
+
 def last_boxed_only_string(string: str) -> Optional[str]:
     """Extract the last LaTeX boxed expression from a string.
 

verl/utils/reward_score/prime_math/__init__.py

@@ -28,9 +28,10 @@
 from pylatexenc import latex2text
 from sympy.parsing import sympy_parser
 
+from verl.utils.py_functional import timeout_limit
+
 from . import math_normalize
 from .grader import math_equal
-from verl.utils.py_functional import timeout_limit
 
 # import math_normalize
 # from grader import math_equal

verl/utils/reward_score/prime_math/grader.py

@@ -95,7 +95,6 @@
 import contextlib
 import math
 import re
-import signal
 from math import isclose
 from typing import Union
 

verl/workers/megatron_workers.py

@@ -262,6 +262,7 @@ def _build_rollout(self, trust_remote_code=False):
             log_gpu_memory_usage("After building sharding manager", logger=logger)
         elif self.config.rollout.name == 'sglang':
             from verl.workers.rollout.sglang_rollout import SGLangRollout
+
             # NOTE(linjunrong): Due to recent fp8 support in SGLang. Now importing any symbol relate to SGLang's model_runner would check CUDA device capability.
             # However, due to veRL's setting, the main process of ray can not find any CUDA device, which would potentially lead to:
             # "RuntimeError: No CUDA GPUs are available".