Build with uv

.python-version

@@ -0,0 +1 @@
+3.10

README.md

@@ -54,26 +54,33 @@ rLLM is an open-source project to fully democratize reinforcement learning (RL)
 
 ## Getting Started 🎯
 ### Installation
-```bash
-# Installing Python 3.10 Environment.
-conda create -n rllm python=3.10 -y
-conda activate rllm
 
-# Installing RLLM dependencies.
+First, make sure [uv](https://docs.astral.sh/uv/getting-started/installation/) is installed.
+
+Then, clone the repository and install dependencies with `uv`:
+
+```bash
+git clone [email protected]:agentica-project/rllm.git
 cd rllm
-pip install -e ./verl
-pip install -e .
+uv sync --extra build
+uv sync --extra build --extra compile
+```
+
+Next, log into wandb:
+
+```bash
+uv run wandb login
 ```
 
 ### Data
 Our raw training data is in `rllm/data/[train|test]/[code|math]/`, along with preprocessing scripts in `rllm/data/preprocess`. To convert the raw data into Parquet files for training, run:
 
 ```bash
 # Download datasets from GDrive, populates rllm/data/[train|test]/[math|code]/*.json
-python scripts/data/download_datasets.py
+uv run scripts/data/download_datasets.py
 
 # Generate parquet files for Deepcoder/DeepscaleR in data/*.parquet
-python scripts/data/[deepcoder|deepscaler]_dataset.py
+uv run scripts/data/[deepcoder|deepscaler]_dataset.py
 ```
 
 ### Training Scripts

pyproject.toml

@@ -0,0 +1,57 @@
+[project]
+name = "rllm"
+version = "0.0.0"
+description = "Distributed Post-Training RL Library for LLMs"
+authors = [{ name = "Agentica Team", email = "[email protected]" }]
+readme = "README.md"
+requires-python = ">=3.10"
+dependencies = [
+    'torch==2.4.0',
+    'torchaudio',
+    'torchvision',
+    'hydra-core==1.3.2',
+    'deepspeed',
+    'dm-tree',
+    'gdown',
+    'google-cloud-aiplatform',
+    'latex2sympy2',
+    'pybind11',
+    'pylatexenc',
+    'pytest',
+    'sentence_transformers',
+    'sortedcontainers',
+    'tabulate',
+    'torchmetrics',
+    'gym',
+    'selenium',
+    'PyMuPDF',
+    'nltk',
+    'browsergym',
+    'gradio',
+    'fire',
+    'vertexai',
+    'e2b_code_interpreter',
+    'firecrawl',
+    'wandb',
+    'verl',
+    'datasets==3.5.0'
+]
+
+[project.optional-dependencies]
+build = ["torch", "setuptools", "packaging"]
+compile = ["flash-attn"]
+
+[tool.uv]
+no-build-isolation-package = ["flash-attn"]
+
+[tool.uv.sources]
+verl = { path = "verl", editable = true }
+
+[[tool.uv.dependency-metadata]]
+name = "flash-attn"
+version = "2.7.4.post1"
+requires-dist = ["torch", "einops"]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"

scripts/deepscaler/train/deepscaler_1.5b_16k.sh

@@ -24,7 +24,7 @@ if [ -z "$MODEL_PATH" ]; then
 fi
 
 # Train over 4 nodes, 8 A100-80GB GPUs per node.
-python3 -m verl.trainer.main_ppo \
+uv run -m verl.trainer.main_ppo \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/rllm/data/deepscaler_train.parquet \
     data.val_files=$HOME/rllm/data/aime.parquet \

scripts/deepscaler/train/deepscaler_1.5b_24k.sh

@@ -24,7 +24,7 @@ if [ -z "$MODEL_PATH" ]; then
 fi
 
 # Train over 4 nodes, 8 A100-80GB GPUs per node.
-python3 -m verl.trainer.main_ppo \
+uv run -m verl.trainer.main_ppo \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/rllm/data/deepscaler_train.parquet \
     data.val_files=$HOME/rllm/data/aime.parquet \

scripts/deepscaler/train/deepscaler_1.5b_8k.sh

@@ -24,7 +24,7 @@ if [ -z "$MODEL_PATH" ]; then
 fi
 
 # Train over a single node, 8 A100-80GB GPUs.
-python3 -m verl.trainer.main_ppo \
+uv run -m verl.trainer.main_ppo \
     algorithm.adv_estimator=grpo \
     data.train_files=$HOME/rllm/data/deepscaler_train.parquet \
     data.val_files=$HOME/rllm/data/aime.parquet \