policy-gradient/GRPO-Zero

• hightopics#1

  Add relevant topics to improve categorization

  Why:

  COPY-PASTE FIX

  Copy

  llm, reinforcement-learning, policy-gradient, deepseek, grpo, low-memory, single-gpu, pytorch, from-scratch

  OpenDoneDismiss
• highreadme#2

  Reposition README opening to highlight lightweight, efficient LLM RL training

  Why:

  CURRENT

  # GRPO:Zero
  
  GRPO training with minimal dependencies (and low GPU memory usage!). We implement almost everything from scratch and only depend on `tokenizers` for tokenization and `pytorch` for training. 
  - No `transformers` and `vLLM` dependencies! 
  - The default config is set to run on a single A40 GPU (48GB VRAM) for a few hours to get good results. (An A40 costs `$0.44` per hour if you rent it from RunPod.)
  - We also support training with a 24GB VRAM GPU (e.g., an RTX 4090 GPU) by offloading the optimizer to CPU. Fortunately, this only adds a small overhead to the training because we only update the policy network a few hundred times during the entire training process.

  COPY-PASTE FIX

  Copy

  # GRPO:Zero
  
  **Train Large Language Models (LLMs) with Group Relative Policy Optimization (GRPO) from scratch, designed for minimal dependencies and low GPU memory usage.** This repository provides an efficient, pure PyTorch implementation of DeepSeek R1's GRPO algorithm, specifically optimized for single-GPU setups (including 24GB VRAM GPUs like the RTX 4090) and completely free of `transformers` and `vLLM` dependencies. Ideal for researchers and practitioners seeking a lightweight, high-performance solution for LLM reinforcement learning.

  OpenDoneDismiss
• mediumhomepage#3

  Add a homepage URL

  Why:

  COPY-PASTE FIX

  Copy

  https://github.com/policy-gradient/GRPO-Zero

  OpenDoneDismiss

Category GEO backends resolved for this scan: google/gemini-2.5-flash, deepseek/deepseek-v4-flash

• CATEGORY QUERY

  What are good options for training LLMs using reinforcement learning without heavy transformer dependencies?

  you: not recommended
  AI recommended (in order):

  1. RLlib (ray-project/ray)
  2. Stable Baselines3 (DLR-RM/stable-baselines3)
  3. Hugging Face Transformers (huggingface/transformers)
  4. PyTorch (pytorch/pytorch)
  5. TensorFlow (tensorflow/tensorflow)
  6. Acme (deepmind/acme)

  AI recommended 6 alternatives but never named policy-gradient/GRPO-Zero. This is the gap to close.

  Show full AI answer
• CATEGORY QUERY

  How can I efficiently train large language models with policy gradients on a single 24GB GPU?

  you: not recommended
  AI recommended (in order):

  1. Hugging Face Transformers
  2. PEFT
  3. LoRA
  4. QLoRA
  5. DeepSpeed ZeRO-2/3
  6. FlashAttention-2
  7. PyTorch FSDP
  8. bitsandbytes
  9. Axolotl
  10. TRL
  11. OpenRLHF

  AI recommended 11 alternatives but never named policy-gradient/GRPO-Zero. This is the gap to close.

  Show full AI answer

• Metadata completeness

  warn

  Suggestion:

• README presence

  pass

• Compared to common alternatives in this category, what is the core differentiator of policy-gradient/GRPO-Zero?

  pass
  ✓AI named policy-gradient/GRPO-Zero explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• If a team adopts policy-gradient/GRPO-Zero in production, what risks or prerequisites should they evaluate first?

  pass
  ✓AI named policy-gradient/GRPO-Zero explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• In one sentence, what problem does the repo policy-gradient/GRPO-Zero solve, and who is the primary audience?

  pass
  ✓AI named policy-gradient/GRPO-Zero explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?