Skip to content

Question about Training Time & [Better] Multi-node Training Scripts. #33

@Walnutes

Description

@Walnutes

Hi, thank you for your fantastic work! I have two questions regarding the training setup

  1. Training Time per Episode
    Could you share the approximate training time per episode (or total training time cost of all 10 episode) on your hardware setup (e.g. NX8 A100)?
  2. [Maybe Better] Multi-Node/Multi-GPU Training Configuration
    I noticed different parallel training parameters in the notion documentation and examples scripts, and I’d like to know the optimal configuration for large-scale training. Could you provide some guidance on how to set the key parameters (e.g. actor_num_nodes, vllm_num_engines, vllm_tensor_parallel_size, vllm_gpu_memory_utilization and (micro_)?(train|rollout)_batch_size) for best performance on specific hardware setup?

Thank you in advance for your time and insights! Looking forward to your reply.

Best regards.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions