inference-server

von huggingface

Starte und teste den prime-rl Inference-Server. Verwende dies, wenn du aufgefordert wirst, eine Inferenz auszuführen, vLLM zu starten, ein Modell zu testen oder den Inference-Server zu starten.

npx skills add https://github.com/huggingface/prime-rl --skill inference-server

ZIP herunterladen GitHub

Inference Server

Starting the server

Always use the inference entry point — never vllm serve or python -m vllm.entrypoints.openai.api_server directly. The entry point runs setup_vllm_env() which configures environment variables (LoRA, multiprocessing) before vLLM is imported.

# With a TOML config
uv run inference @ path/to/config.toml

# With CLI overrides
uv run inference --model.name Qwen/Qwen3-0.6B --model.max_model_len 2048 --model.enforce_eager

# Combined
uv run inference @ path/to/config.toml --server.port 8001 --gpu-memory-utilization 0.5

SLURM scheduling

The inference entrypoint supports optional SLURM scheduling, following the same patterns as SFT and RL.

Single-node SLURM

# inference_slurm.toml
output_dir = "/shared/outputs/my-inference"

[model]
name = "Qwen/Qwen3-8B"

[parallel]
tp = 8

[slurm]
job_name = "my-inference"
partition = "cluster"

uv run inference @ inference_slurm.toml

Multi-node SLURM (independent vLLM replicas)

Each node runs an independent vLLM instance. No cross-node parallelism — TP and DP must fit within a single node's GPUs.

# inference_multinode.toml
output_dir = "/shared/outputs/my-inference"

[model]
name = "PrimeIntellect/INTELLECT-3-RL-600"

[parallel]
tp = 8
dp = 1

[deployment]
type = "multi_node"
num_nodes = 4
gpus_per_node = 8

[slurm]
job_name = "my-inference"
partition = "cluster"

Dry run

Add dry_run = true to generate the sbatch script without submitting:

uv run inference @ config.toml --dry-run true

Custom endpoints

The server extends vLLM with:

/v1/chat/completions/tokens — accepts token IDs as prompt input (used by multi-turn RL rollouts)
/update_weights — hot-reload model weights from the trainer
/load_lora_adapter — load LoRA adapters at runtime
/init_broadcaster — initialize weight broadcast for distributed training

Testing the server

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-0.6B",
    "messages": [{"role": "user", "content": "Hi"}],
    "max_tokens": 50
  }'

Key files

src/prime_rl/entrypoints/inference.py — entrypoint with local/SLURM routing
src/prime_rl/inference/server.py — vLLM env setup
src/prime_rl/configs/inference.py — InferenceConfig and all sub-configs
src/prime_rl/inference/vllm/server.py — FastAPI routes and vLLM monkey-patches
src/prime_rl/templates/inference.sbatch.j2 — SLURM template (handles both single and multi-node)
configs/debug/infer.toml — minimal debug config