// How to install prime-rl and its optional dependencies. Use when setting up the project, installing extras like deep-gemm for FP8 models, or troubleshooting dependency issues.
How the prime-rl config system works — TOML files, CLI overrides, composition, and special patterns. Use when creating configs, debugging config errors, or overriding values via CLI.
How to prepare and publish GitHub releases for prime-rl. Use when drafting release notes, tagging versions, or publishing releases.
Monitor an ongoing prime-rl training run — find the output directory, tail logs, check key metrics, inspect SLURM jobs, and restart safely. Use when asked to check on a run, debug training, or investigate performance.
Launch and monitor prime-rl training runs. Use when starting, supervising, or debugging an RL/SFT run. Routes to `start-run` (entrypoints + how to launch) and `monitor-run` (logs, metrics, check-ins).
How to launch prime-rl training runs — the `rl`, `sft`, and `inference` entrypoints, their config classes, and single-node/SLURM/dry-run modes. Use when starting a run or picking the right entrypoint.
| name | install |
| description | How to install prime-rl and its optional dependencies. Use when setting up the project, installing extras like deep-gemm for FP8 models, or troubleshooting dependency issues. |
prime-rl is a monorepo with submodules. Use the install script when bootstrapping a fresh machine:
bash scripts/install.sh # clones, inits submodules, installs uv, runs `uv sync --all-extras`
For an existing clone, init submodules explicitly:
git submodule update --init -- deps/verifiers deps/renderers deps/research-environments deps/pydantic-config
Do not run git submodule update --init --recursive without paths — it tries to clone the private configs/private submodule and aborts for users without access. scripts/install.sh walks submodules one at a time and skips failures, so it works for everyone.
uv sync # core only
uv sync --group dev # + pytest, ruff, pre-commit
uv sync --all-extras # recommended: envs, flash-attn, flash-attn-cute, etc.
The envs extra installs every env workspace listed in [tool.uv.workspace]. Adding a new env means adding it to members, the envs extra, and [tool.uv.sources].
When bumping a package past the workspace-wide exclude-newer = "7 days" window, add it (and any newly-required transitives) to [tool.uv.exclude-newer-package] before refreshing uv.lock.
CUDA_HOME=/usr/local/cuda uv pip install mamba-ssm
Requires nvcc. Without mamba-ssm, NemotronH falls back to HF's pure-PyTorch SSD path, which computes softplus in bf16 and yields ~0.4 KL divergence vs vLLM. Do not install causal-conv1d unless your GPU arch matches the prebuilt kernels — the code falls back to nn.Conv1d when it's absent.
uv sync --group fp8-inference # installs the prebuilt deep-gemm wheel
scripts/install_ep_kernels.sh auto-detects the CUDA toolkit matching torch and the GPU arch, builds NVSHMEM + DeepEP from source, and skips if deep_ep already imports.
bash scripts/install_ep_kernels.sh
Flags: --workspace DIR, --deepep-ref REF (default 73b6ea4), --nvshmem-ver VER (default 3.3.24), --configure-drivers (multi-node IBGDA; needs sudo + reboot).
Verify: uv run python -c 'import deep_ep; print(deep_ep.__file__)'.
pyproject.toml — dependencies, extras, dependency groupsuv.lock — pinned lockfile (refresh with uv sync --all-extras)scripts/install.sh — bootstrap installerscripts/install_ep_kernels.sh — DeepEP build script