skills.rest
ricable
ricableCOMMUNITY
Explore all skills by this author →
GitHub

torchforge-rl-training

Coordinate distributed PyTorch-native RL training for agentic language models.

Updated Feb 15, 2026
Search Tags:#reinforcement learning#pytorch#distributed training#vllm#grpo#torchtitan#monarch
One-click install
npx skills add https://github.com/ricable/mcai --skill torchforge-rl-training-ricable
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: torchforge-rl-training
Source: https://github.com/ricable/mcai/tree/main/.agents/skills/torchforge-rl-training
Command: npx skills add https://github.com/ricable/mcai --skill torchforge-rl-training-ricable

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

This Skill helps you design, configure, and run agentic reinforcement learning workflows without mixing algorithm logic into distributed infrastructure concerns.

Core Features & Use Cases

  • Algorithm experimentation: Implement or adapt RL losses such as GRPO, DAPO, SAPO, and REINFORCE with clean PyTorch-native abstractions.
  • Distributed training orchestration: Coordinate trainer, generator, and reference model services across multiple GPUs or nodes using Monarch and TorchTitan.
  • Practical use case: Use it to launch GRPO training for math reasoning, tune reward functions, or scale an SFT pipeline while keeping inference, training, and weight sync separated.

Quick Start

Use torchforge to train a model by selecting the right YAML config, defining a reward or loss function, and launching the corresponding SFT or GRPO entry point.

Frequently Asked Questions about torchforge-rl-training

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I scale reinforcement learning training across multiple GPUs without mixing algorithm logic into infrastructure code?▼

Scale reinforcement learning training by coordinating trainer, generator, and reference model services across multiple GPUs using Monarch, TorchTitan, and vLLM. This approach separates algorithm logic from distributed infrastructure concerns, allowing you to configure workflows via YAML.

Can I use vLLM for inference during distributed PyTorch GRPO training?▼

Yes, you can use vLLM for inference during distributed PyTorch GRPO training. The workflow requires vLLM inference with weight synchronization to coordinate generation alongside the trainer and reference model services across nodes.

How do I implement custom RL losses like GRPO or DAPO for agentic language models?▼

Implement custom RL losses like GRPO, DAPO, and REINFORCE using clean PyTorch-native abstractions. This allows you to experiment with algorithm logic independently from the distributed training orchestration and generation infrastructure.

What is the best way to orchestrate trainer, generator, and reference models for agentic workflows?▼

Orchestrate trainer, generator, and reference models using async Monarch services and TorchTitan-backed model parallelism. This architecture separates inference, training, and weight sync, enabling scalable configuration-driven deployment for agentic workflows.

Does this workflow support SFT pipelines alongside reinforcement learning experimentation?▼

Yes, the workflow supports SFT pipelines alongside reinforcement learning experimentation. You can launch GRPO training for math reasoning, tune reward functions, or scale an SFT pipeline while keeping inference, training, and weight sync separated.

Why do I need async Monarch services and TorchTitan for distributed RL training?▼

Async Monarch services and TorchTitan are required for distributed RL training to coordinate separate trainer, generator, and reference model services across multiple GPUs. TorchTitan provides model parallelism while Monarch handles the asynchronous service orchestration.

Related Skills

Explore More Software Engineering Skills→
SignmanalSignmanal
1

huggingface-accelerate

Eliminate hardware-specific boilerplate for distributed PyTorch model training.

Community
Intermediate
zeronx798zeronx798

evaluating-llms-harness

Benchmark large language models across academic and reasoning datasets.

Community
Advanced
tianhao909tianhao909
1

torchforge-rl-training

Train PyTorch-native RL agents with GRPO, DAPO, and SAPO algorithms.

Community
Advanced
jackspacejackspace
16

pufferlib

Optimize high-throughput PPO+LSTM reinforcement learning training with vectorized environments.

Community
Advanced