|
1 | | -# Installation |
| 1 | +# Installation Guide |
2 | 2 |
|
3 | | -Claw-R1 uses [veRL](https://github.com/volcengine/verl) as its training backend and [Ray](https://docs.ray.io/) for distributed execution. |
| 3 | +Claw-R1 uses the same environment setup as `verl`. |
4 | 4 |
|
5 | | -## Prerequisites |
| 5 | +## Base Environment |
6 | 6 |
|
7 | | -| Requirement | Version | Notes | |
8 | | -|---|---|---| |
9 | | -| Python | 3.10+ | Conda recommended | |
10 | | -| CUDA | 11.8+ | Required for GPU training | |
11 | | -| Conda / Mamba | latest | For environment isolation | |
12 | | -| Git | any | For cloning repositories | |
| 7 | +Follow the official [`verl` installation guide](https://verl.readthedocs.io/en/latest/start/install.html), but make sure the environment ends up with `verl==0.7.0`. |
13 | 8 |
|
14 | | -## Step 1 — Create the Environment |
| 9 | +If you want a broader overview of the base training workflow, the [`verl` quickstart](https://verl.readthedocs.io/en/latest/start/quickstart.html) is also useful. |
15 | 10 |
|
16 | | -```bash |
17 | | -conda create -n clawr1 python=3.10 -y |
18 | | -conda activate clawr1 |
19 | | -``` |
| 11 | +## What This Means for Claw-R1 |
20 | 12 |
|
21 | | -## Step 2 — Install veRL |
| 13 | +Once the `verl` environment is working, Claw-R1 should run in the same environment. In practice, that means you can: |
22 | 14 |
|
23 | | -Claw-R1 requires the nightly version of veRL installed from source: |
| 15 | +- prepare a Python environment with `verl==0.7.0` |
| 16 | +- clone this repository |
| 17 | +- run Claw-R1 commands directly from the repository root |
24 | 18 |
|
25 | | -```bash |
26 | | -git clone https://github.com/volcengine/verl && cd verl |
27 | | -pip install --no-deps -e . |
28 | | -cd .. |
29 | | -``` |
| 19 | +You do not need to install Claw-R1 as a separate package. |
30 | 20 |
|
31 | | -## Step 3 — Install Claw-R1 |
32 | | - |
33 | | -```bash |
34 | | -git clone https://github.com/AgentR1/Claw-R1 && cd Claw-R1 |
35 | | -pip install -e . |
36 | | -``` |
37 | | - |
38 | | -## Step 4 — Install Ray and FastAPI |
39 | | - |
40 | | -```bash |
41 | | -pip install "ray[default]" fastapi uvicorn |
42 | | -``` |
43 | | - |
44 | | -## Verify Installation |
45 | | - |
46 | | -Run the following checks to ensure everything is installed correctly: |
47 | | - |
48 | | -```bash |
49 | | -# Check veRL |
50 | | -python -c "import verl; print('veRL:', verl.__version__)" |
51 | | - |
52 | | -# Check Ray |
53 | | -python -c "import ray; print('Ray:', ray.__version__)" |
54 | | - |
55 | | -# Check Claw-R1 gateway |
56 | | -python -m claw_r1.gateway.gateway --help |
57 | | -``` |
58 | | - |
59 | | -!!! tip "GPU Memory" |
60 | | - The Gateway Server is CPU-only and lightweight. Rollout workers (vLLM) and Training workers require separate GPU pools. See [Async Training](../components/async-training.md) for multi-GPU configuration. |
61 | | - |
62 | | -## What's Next |
63 | | - |
64 | | -- [Quick Start](quickstart.md) — run your first training loop |
65 | | -- [Configuration Reference](../configuration/index.md) — customize the setup for your hardware |
| 21 | +The documentation in this repository intentionally does not duplicate a separate environment guide, so that the infrastructure setup stays aligned with `verl`. |
0 commit comments