Dependency conflicts: Verl requires tensordict<=0.6.2 but Torch 2.7.0 (CUDA 12.8, B200 GPUs) requires tensordict>=0.7.0

[jatinS-dev]

Description

I am trying to run Agent-R1 on HiPerGator using the B200 GPU partition (hpg-b200). The new B200 GPUs require PyTorch ≥2.7.0 with CUDA 12.8 for sm_100 architecture support. However, Verl and vLLM dependencies were originally pinned for Torch 2.3.0, which creates conflicts.

Steps to Reproduce

1. Create a clean conda env with Python 3.10.
2. Install Verl (0.2.0.dev0) as per project instructions.
3. Install PyTorch stack for B200:

   pip install torch==2.7.0+cu128 torchvision==0.22.0+cu128 torchaudio==2.7.0+cu128 --index-url https://download.pytorch.org/whl/cu128

4. Try installing tensordict, xformers, and vllm.

Actual Results

• verl 0.2.0.dev0 requires tensordict<=0.6.2.
• torch==2.7.0 requires tensordict>=0.7.0.
• vllm 0.4.x (used in Agent-R1) only works with torch==2.3.0.
• vllm >=0.10.x works with Torch 2.7.0, but introduces breaking API changes (model_hf_config removed).

Example error

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed.
verl 0.2.0.dev0 requires tensordict<=0.6.2, which is not installed.
torchaudio 2.7.0+cu128 requires torch==2.7.0, but you have torch 2.7.1 which is incompatible.
torchvision 0.22.0+cu128 requires torch==2.7.0, but you have torch 2.7.1 which is incompatible.

Expected Results

Ability to install and run Agent-R1/Verl with PyTorch 2.7.0 + CUDA 12.8 (for B200 GPUs) without dependency conflicts.

CC

@lyumengxian