Environment:Volcengine Verl Python Core Dependencies
Metadata
| Field | Value |
|---|---|
| Sources | verl|https://github.com/volcengine/verl |
| Domains | Infrastructure |
| Last Updated | 2026-02-07 17:00 GMT |
Overview
Core Python dependencies for the verl RL training framework including PyTorch, transformers, Ray, and tensordict.
Description
verl requires a set of core Python packages. PyTorch >= 2.4 is needed for FSDP2 (fully_shard API). tensordict >= 0.8.0 and <= 0.10.0 (excluding 0.9.0) is required for the DataProto protocol. Ray >= 2.41.0 is needed for distributed orchestration. numpy must be < 2.0.0 for compatibility.
Usage
These are mandatory for all verl workflows.
System Requirements
- Python 3.10+ recommended
- Linux OS
Dependencies
From setup.py:
- torch >= 2.4.0 (required for FSDP2)
- tensordict >= 0.8.0, <= 0.10.0, != 0.9.0
- ray[default] >= 2.41.0
- transformers
- accelerate
- datasets
- peft
- hydra-core
- numpy < 2.0.0
- pyarrow >= 19.0.0
- packaging >= 20.0
- wandb
- pandas
- tensorboard
Credentials
- WANDB_ENTITY: W&B logging entity
- MLFLOW_TRACKING_URI: MLflow tracking URI
- SWANLAB_API_KEY: SwanLab logging API key
- TENSORBOARD_DIR: TensorBoard output directory
Quick Install
pip install verl
Or install dependencies directly:
pip install torch>=2.4.0 transformers accelerate datasets peft hydra-core "numpy<2.0.0" "ray[default]>=2.41.0" "tensordict>=0.8.0,<=0.10.0,!=0.9.0" wandb pandas tensorboard "pyarrow>=19.0.0" "packaging>=20.0"
Code Evidence
From verl/utils/fsdp_utils.py:511:
assert CPUOffloadPolicy is not None, "PyTorch version >= 2.4 is required for using fully_shard API (FSDP2)"
And from verl/protocol.py:567-568:
assert version.parse(tensordict.__version__) >= version.parse("0.10.0")
Common Errors
| Error | Solution |
|---|---|
| "PyTorch version >= 2.4 is required for using fully_shard API (FSDP2)" | Upgrade PyTorch |
| tensordict version assertion | Install correct tensordict version |
| "No module named 'ray'" | pip install "ray[default]>=2.41.0" |
Compatibility Notes
numpy 2.0+ is explicitly excluded. tensordict 0.9.0 has known incompatibilities.