mirrors/open-r1
2
0
Fork 0
mirror of https://github.com/huggingface/open-r1.git synced 2026-06-24 01:54:06 +00:00
Code Issues Projects Releases Packages Wiki Activity
open-r1/slurm/README.md
Guilherme Penedo 7835979801
 adds support for running GRPO on IOI problems (#495) 
* adds support for running GRPO on IOI problems

* nit

* bugfixes + recipe

* added piston info and readme changes

* readme updates

* run isort to fix checks

* Update src/open_r1/rewards.py

Co-authored-by: Edward Beeching <edbeeching@users.noreply.github.com>

* adding ioi test

* fix merge issues with python slow tests

* style

* generalize piston workers

* generalize readme

* fix extract code

* finalize slow tests

---------

Co-authored-by: Edward Beeching <edbeeching@users.noreply.github.com>
Co-authored-by: edbeeching <edbeeching@gmail.com>
2025-03-21 08:48:00 +01:00

938 B
Raw Permalink Blame History

Serving DeepSeek-R1 on 2x8 H100 SLURM nodes with SGLang

  1. Set up the environment (adjust for your cuda version):
conda create -n sglang124 python=3.11
conda activate sglang124

pip install torch==2.5.1 --index-url https://download.pytorch.org/whl/cu124

pip install sgl-kernel --force-reinstall --no-deps
pip install "sglang[all]>=0.4.2.post4" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer/
  1. Run the server and wait for the model to load:
sbatch slurm/serve_r1.slurm -m "/fsx/deepseek-r1-checkpoint" -e "sglang124"
  1. Run the data generation script:
python scripts/generate_reasoning.py \
    --dataset-name "AI-MO/NuminaMath-1.5" \
    --output-file "numinamath_r1_generations.jsonl" \
    --prompt-column "problem" \
    --uuid-column "problem" \
    --api-addr "<SGLANG_SERVER_ADDRESS>:39877" \
    --num-generations 2 \
    --max-tokens 16384 \
    --max-concurrent 200