/post-training-torchforge

name: torchforge-rl-training
description: Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.
version: 1.0.0
author: Orchestra Research
license: MIT
tags: [Reinforcement Learning, PyTorch, GRPO, SFT, Monarch, TorchTitan, Meta]
dependencies: [torch>=2.9.0, torchtitan>=0.2.0, vllm, monarch]

torchforge: PyTorch-Native Agentic RL Library

torchforge is Meta's PyTorch-native RL library that separates infrastructure concerns from algorithm concerns. It enables rapid RL research by letting you focus on algorithms while handling distributed training, inference, and weight sync automatically.

When to Use torchforge

**Choose torchforge when you need:**

• Clean separation between RL algorithms and infrastructure
• PyTorch-native abstractions (no Ray dependency)
• Easy algorithm experimentation (GRPO, DAPO, SAPO in ~100 lines)
• Scalable training with Monarch actor system
• Integration with TorchTitan for model parallelism

**Consider alternatives when:**

• You need production-ready stability → use **miles** or **verl**
• You want Megatron-native training → use **slime**
• torchforge is experimental and APIs may change

Key Features

• **Algorithm isolation**: Implement RL algorithms without touching infrastructure
• **Scalability**: From single GPU to thousands via Monarch
• **Modern stack**: TorchTitan (training), vLLM (inference), TorchStore (sync)
• **Loss functions**: GRPO, DAPO, CISPO, GSPO, SAPO built-in

Architecture Overview

┌─────────────────────────────────────────────────────────┐
│ Application Layer (Your Code)                           │
│ - Define reward models, loss functions, sampling        │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│ Forge API Layer                                         │
│ - Episode, Group dataclasses                           │
│ - Service interfaces (async/await)                      │
└─────────────────────┬───────────────────────────────────┘
                      │
┌─────────────────────▼───────────────────────────────────┐
│ Distributed Services (Monarch)                          │
│ ├── Trainer (TorchTitan FSDP)                          │
│ ├── Generator (vLLM inference)                          │
│ ├── Reference Model (frozen KL baseline)               │
│ └── Reward Actors (compute rewards)                    │
└─────────────────────────────────────────────────────────┘

Installation

# Create environment
conda create -n forge python=3.12
conda activate forge

# Install (handles PyTorch nightly + dependencies)
./scripts/install.sh

# Verify
python -c "import torch, forge, vllm; print('OK')"

ROCm Installation

./scripts/install_rocm.sh

Quick Start

SFT Training (2+ GPUs)

python -m apps.sft.main --config apps/sft/llama3_8b.yaml

GRPO Training (3+ GPUs)

python -m apps.grpo.main --config apps/grpo/qwen3_1_7b.yaml

---

Workflow 1: GRPO Training for Math Reasoning

Use this workflow for training reasoning models with group-relative advantages.

Prerequisites Checklist

• [ ] 3+ GPUs (GPU0: trainer, GPU1: ref_model, GPU2: generator)
• [ ] Model from HuggingFace Hub
• [ ] Training dataset (GSM8K, MATH, etc.)

Step 1: Create Configuration

# config/grpo_math.yaml
model: "Qwen/Qwen2.5-7B-Instruct"

dataset:
  path: "openai/gsm8k"
  split: "train"
  streaming: true

training:
  batch_size: 4
  learning_rate: 1e-6
  seq_len: 4096
  dtype: bfloat16
  gradient_accumulation_steps: 4

grpo:
  n_samples: 8           # Responses per prompt
  clip_low: 0.2
  clip_high: 0.28
  beta: 0.1              # KL penalty coefficient
  temperature: 0.7

services:
  generator:
    procs: 1
    num_replicas: 1
    with_gpus: true
  trainer:
    procs: 1
    num_replicas: 1
    with_gpus: true
  ref_model:
    procs: 1
    num_replicas: 1
    with_gpus: true

Step 2: Define Reward Function

# rewards.py
# Reward functions are in forge.data.rewards
from forge.data.rewards import MathReward, ThinkingReward
import re

# Or define your own reward function
class CustomMathReward:
    def __call__(self, prompt: str, response: str, target: str) -> float:
        # Extract answer from response
        match = re.search(r'\\boxed{([^}]+)}', response)
        if not match:
            return 0.0

        answer = match.group(1).strip()
        return 1.0 if answer == target else 0.0

Step 3: Launch Training

python -m apps.grpo.main --config config/grpo_math.yaml

Step 4: Monitor Progress

• [ ] Check W&B dashboard for loss curves
• [ ] Verify entropy is decreasing (policy becoming more deterministic)
• [ ] Monitor KL divergence (should stay bounded)

---

Workflow 2: Custom Loss Function

Use this workflow to implement new RL algorithms.

Step 1: Create Loss Class

# src/forge/losses/custom_loss.py
import torch
import torch.nn as nn

class CustomLoss(nn.Module):
    def __init__(self, clip_range: float = 0.2, beta: float = 0.1):
        super().__init__()
        self.clip_range = clip_range
        self.beta = beta

    def forward(
        self,
        logprobs: torch.Tensor,
        ref_logprobs: torch.Tensor,
        advantages: torch.Tensor,
        padding_mask: torch.Tensor,
    ) -> torch.Tensor:
        # Compute importance ratio
        ratio = torch.exp(logprobs - ref_logprobs)

        # Clipped policy gradient
        clipped_ratio = torch.clamp(
            ratio,
            1 - self.clip_range,
            1 + self.clip_range
        )
        pg_loss = -torch.min(ratio * advantages, clipped_ratio * advantages)

        # KL penalty
        kl = ref_logprobs - logprobs

        # Apply mask and aggregate
        masked