KO
|
EN
gitlite — search
Search
#javascript
#python
#hacktoberfest
#react
#ai
#typescript
#llm
#go
#golang
#android
#machine-learning
#rust
#deep-learning
#linux
modal-grpo
★ 19
Open GitHub ↗
No description available.
Download README (.md)
Explore Similar Repositories
Alpha-Advanced
:
Mod for Minecraft Alpha 1.1.2_01
Scalable-Capital-Transactions-Exporter
:
Scalable Capital Transactions CSV Exporter for Portfolio Performance
wretch
:
This is my fetch script written in Rust, as one of my first projects
zotero-ai-tab
:
AI sidebar plugin for Zotero, compatible with OpenAI, supports GPT, DeepSeek, Qwen, Doubao and other LLMs.
nmea2000
:
NMEA 2000 Python encoder and decoder based on canboat database
// repository documentation
Was this content helpful?
★ 0
(0 ratings)
Select Rating:
★
★
★
★
★
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
modal-grpo
?
Download (.md)
# GRPO Training Examples This a repository that shows how to run a simple GRPO example. There are two main implementations here in this repository 1. `grpo_simple.py` : This shows how you can use a single Modal A100 to run 2 GRPO policy steps at each point 2. `grpo.ipynb` : This is a notebook implementation that can be ran on any GPU setup. I'm trying to figure out how can I use the GRPO trainer to do distributed multi-node training because there doesn't seem to be a way to do so? Also the `use_vllm` integration keeps causing me to go OOM for some reason and that's a bit weird. ## Current Limitations - Multi-GPU support is currently not working as expected due to device mapping issues - The implementations are currently optimized for single GPU usage - The Modal implementation is configured for A100 GPUs specifically ## Contributing Contributions are welcome! If you can help solve the multi-GPU device mapping issues or have other improvements, please feel free to submit a PR. Some specific areas that need attention: - Fixing device mapping for multi-GPU training - Optimizing batch sizes and memory usage - Adding support for different model architectures ## License [Insert your chosen license here]
