cyberneticlibrary

Train agents with TorchForge

torchforge-rl-trainingskillsetup L49,423
Orchestra-Research/AI-Research-SKILLs
What it does

Optimize RL training with PyTorch-native abstractions

Best for

PyTorch-native RL training with hardware acceleration and custom loss functions.

Inputs
  • · Policy model
  • · Reward signal
  • · Experience replay buffer
  • · Optimizer config
Outputs
  • · Optimized policy
  • · Training logs
  • · Performance metrics
Requires
  • · torchforge
  • · torch>=2.0
  • · ray (optional)
Preconditions
  • · PyTorch model defined
  • · Reward function computable
  • · GPU available
Failure modes
  • · Gradient explosion
  • · Dead gradients
  • · Experience replay stale
  • · Batch too large
Trust signals
  • · PyTorch ecosystem
  • · Direct tensor operations