| # Real-search GRPO training & evaluation (tested in conda env: genrl, Python 3.11) | |
| # Install PyTorch for your CUDA version first, then: | |
| # pip install -r requirements.txt | |
| # Optional (recommended on H100): pip install flash-attn --no-build-isolation | |
| # Core training | |
| ray==2.55.1 | |
| sglang==0.5.10.post1 | |
| transformers==5.3.0 | |
| torch==2.9.1 | |
| torchvision==0.24.1 | |
| torchaudio==2.9.1 | |
| triton==3.5.1 | |
| accelerate==1.13.0 | |
| peft==0.19.1 | |
| deepspeed==0.19.0 | |
| # Config & hub | |
| hydra-core==1.3.2 | |
| omegaconf==2.3.0 | |
| huggingface_hub==1.14.0 | |
| safetensors==0.7.0 | |
| tokenizers==0.22.2 | |
| datasets==4.8.5 | |
| # Eval & tools | |
| openai==2.6.1 | |
| pydantic==2.12.5 | |
| requests==2.32.3 | |
| httpx==0.28.1 | |
| aiohttp==3.13.5 | |
| fastapi==0.136.1 | |
| uvicorn==0.46.0 | |
| tiktoken==0.12.0 | |
| sentencepiece==0.2.1 | |
| # Utilities | |
| numpy==1.26.4 | |
| pandas==3.0.3 | |
| pyarrow==24.0.0 | |
| fsspec==2026.2.0 | |
| einops==0.8.2 | |
| tqdm==4.67.3 | |
| PyYAML==6.0.1 | |
| jsonschema==4.26.0 | |
| codetiming==1.4.0 | |
| wandb==0.27.0 | |
| tensorboard==2.20.0 | |
| psutil==7.2.2 | |
| Jinja2==3.1.6 | |
| rich==15.0.0 | |
| # Optional acceleration (install separately if needed) | |
| # flash-attn==4.0.0b13 | |