[install] chore: add pyproject.toml. make vllm default dependency (#63)

docs/start/quickstart.rst

@@ -18,7 +18,7 @@ Prerequisite:
 
 - the latest version of ``verl`` and its dependencies installed following the installation guide. Using the docker image is recommended.
 
-- an GPU with at least 32 GB memory
+- an GPU with at least 24 GB HBM
 
 
 Dataset Introduction
@@ -108,7 +108,7 @@ Set the ``data.train_files`` ,\ ``data.val_files``, ``actor_rollout_ref.model.pa
     trainer.nnodes=1 \
     trainer.save_freq=10 \
     trainer.test_freq=10 \
-    trainer.total_epochs=15 $@ 2>&1 | tee verl_demo.log
+    trainer.total_epochs=15 2>&1 | tee verl_demo.log
 
 You are expected to see the following logs, indicating training in progress. The key metric ``val/test_score/openai/gsm8k`` is computed every ``trainer.test_freq`` steps:
 
@@ -129,15 +129,14 @@ To enable ``wandb`` for experiment tracking, set the following configs:
     trainer.project_name=$YOUR_PROJECT_NAME \
     trainer.experiment_name=$YOUR_RUN_NAME \
 
-If you encounter out of memory issues, enable the following configs would help:
+If you encounter out of memory issues with HBM less than 32GB, enable the following configs would help:
 
-- actor_rollout_ref.actor.ppo_micro_batch_size=1 \
-
-- critic.ppo_micro_batch_size=1 \
-
-- actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
+.. code-block:: bash
 
-- critic.model.fsdp_config.optimizer_offload=False \
+    actor_rollout_ref.actor.ppo_micro_batch_size=1 \
+    critic.ppo_micro_batch_size=1 \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    critic.model.fsdp_config.optimizer_offload=True \
 
 For the full set of configs, please refer to :ref:`config-explain-page` for detailed explaination and performance tuning.
 

pyproject.toml

@@ -0,0 +1,78 @@
+# -------------------------------
+# build-system
+# -------------------------------
+[build-system]
+requires = [
+    "setuptools>=61.0",
+    "wheel"
+]
+build-backend = "setuptools.build_meta"
+
+# -------------------------------
+# project (PEP 621 metadata)
+# -------------------------------
+[project]
+name = "verl"
+# We'll mark the version as "dynamic" because it's read from the file "verl/version/version" 
+# (PEP 621 calls this "dynamic version"). 
+# The actual version is specified in the [tool.setuptools.dynamic] section below.
+dynamic = ["version"]
+
+description = "veRL: Volcano Engine Reinforcement Learning for LLM"
+license = {file = "LICENSE"}  # or "Apache-2.0", if you prefer an SPDX identifier
+readme = {file = "README.md", content-type = "text/markdown"}
+requires-python = ">=3.8"
+
+authors = [
+  { name = "Bytedance - Seed - MLSys", email = "[email protected]" },
+  { name = "Bytedance - Seed - MLSys", email = "[email protected]" },
+]
+
+# Dependencies corresponding to install_requires in setup.py
+dependencies = [
+    "accelerate",
+    "codetiming",
+    "datasets",
+    "dill",
+    "hydra-core",
+    "numpy",
+    "pybind11",
+    "ray==2.10",
+    "tensordict",
+    "transformers",
+    "vllm<=0.6.3",
+]
+
+# Optional dependencies (extras_require in setup.py)
+[project.optional-dependencies]
+test = [
+  "pytest", "yapf"
+]
+
+# URLs
+[project.urls]
+Homepage = "https://github.com/volcengine/verl"
+
+# -------------------------------
+# tool.setuptools - Additional config
+# -------------------------------
+[tool.setuptools]
+# True means `setuptools` will attempt to include all relevant files in package_data automatically.
+# 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'
+[tool.setuptools.dynamic]
+version = {file = "verl/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 = [
+  "version/*",
+  "trainer/config/*.yaml"
+]

requirements.txt

@@ -1,5 +1,3 @@
-# vllm==0.6.3 # vllm is installed in image building to avoid ray conflicts
-# TODO: add version info to requirements
 accelerate
 codetiming
 datasets
@@ -8,5 +6,6 @@ hydra-core
 numpy
 pybind11
 ray==2.10
-tensordict < 0.6
+tensordict<0.6
 transformers
+vllm<=0.6.3

setup.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+# setup.py is the fallback installation script when pyproject.toml does not work
 from setuptools import setup, find_packages
 import os
 
@@ -25,14 +26,8 @@
     required = f.read().splitlines()
     install_requires = [item.strip() for item in required if item.strip()[0] != '#']
 
-install_optional = [
-    'vllm==0.6.3',
-    'torch==2.4.0', # required by vllm
-]
-
 extras_require = {
-    'single-controller': ['ray', 'kubernetes'],
-    'test': ['pytest']
+    'test': ['pytest', 'yapf']
 }
 
 from pathlib import Path
@@ -56,4 +51,4 @@
     include_package_data=True,
     long_description=long_description,
     long_description_content_type='text/markdown'
-)
+)