Deploy MiniCPM5-1B and MiniCPM5-2B with SGLang

OpenAI-compatible server with RadixAttention prefix cache. Best fit for tool calling, batched eval pipelines, and high-concurrency serving.

Required input

Var	Example	Default
`MODEL_PATH`	`openbmb/MiniCPM5-2B` or `openbmb/MiniCPM5-1B`	required;
`PORT`	`30000`	`30000`
`GPU_ID`	`0`	`0`
`CTX_LEN`	`131072` (128 K)	`131072`
`MEM_FRAC`	`0.85`	`0.85`
`TOOL_PARSER`	`minicpm5`	`minicpm5`; use `auto` if you want template detection

Steps

1. Install (once)

pip install "sglang[srt]>=0.5.16"          # latest, requires CUDA 13.x driver
# pip install "sglang==0.5.6.post3"        # fallback for CUDA 12.x driver hosts

For tool calling, install from main — the minicpm5 parser (PR #25600, merged 2026-05-22) is not in any pip release yet (v0.5.12.post1 was branched earlier). Plain chat works on the pip release; only --tool-call-parser minicpm5 needs main:

pip install "git+https://github.com/sgl-project/sglang.git@main#subdirectory=python"

2. Recommended runtime env vars

export VLLM_WORKER_MULTIPROC_METHOD=spawn
export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
export SGLANG_DISABLE_CUDNN_CHECK=1

3. Launch

CUDA_VISIBLE_DEVICES=${GPU_ID} python -m sglang.launch_server \
    --model-path "${MODEL_PATH}" \
    --served-model-name MiniCPM5-2B \
    --dtype bfloat16 \
    --context-length ${CTX_LEN} \
    --mem-fraction-static ${MEM_FRAC} \
    --tool-call-parser ${TOOL_PARSER} \
    --host 0.0.0.0 \
    --port ${PORT}

Wait for The server is fired up and ready to roll! .

Speculative decoding (DSpark): we also release MiniCPM5-2B-DSpark, a DSpark draft model trained for MiniCPM5-2B. Enable it in SGLang to accelerate decoding while keeping the target model's outputs unchanged:

CUDA_VISIBLE_DEVICES=${GPU_ID} python -m sglang.launch_server \
    --model-path "${MODEL_PATH}" \
    --served-model-name MiniCPM5-2B \
    --dtype bfloat16 \
    --context-length ${CTX_LEN} \
    --mem-fraction-static ${MEM_FRAC} \
    --tool-call-parser ${TOOL_PARSER} \
    --host 0.0.0.0 \
    --port ${PORT} \
    --speculative-algorithm DSPARK \
    --speculative-draft-model-path openbmb/MiniCPM5-2B-DSpark \
    --speculative-dspark-block-size 7

4. Validate

curl http://localhost:${PORT}/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "MiniCPM5-2B",
        "messages": [{"role": "user", "content": "1+1=?"}],
        "temperature": 1.0, "top_p": 0.95, "max_tokens": 64,
        "chat_template_kwargs": {"enable_thinking": true}
    }'

Expected: choices[0].message.content contains "2".

Tool calling

SGLang includes a native MiniCPM5 XML tool-call parser. Keep TOOL_PARSER=minicpm5 (or auto) when launching, then send an OpenAI-style tools request:

curl http://localhost:${PORT}/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "MiniCPM5-2B",
        "messages": [{"role": "user", "content": "What is the weather in Beijing?"}],
        "tools": [{
            "type": "function",
            "function": {
                "name": "get_weather",
                "description": "Get current weather for a city",
                "parameters": {
                    "type": "object",
                    "properties": {"city": {"type": "string"}},
                    "required": ["city"]
                }
            }
        }],
        "tool_choice": "auto",
        "temperature": 1.0, "top_p": 0.95,
        "max_tokens": 256
    }'

Offline / batched (Engine API)

import sglang as sgl

llm = sgl.Engine(model_path="${MODEL_PATH}", tp_size=1, mem_fraction_static=0.8, context_length=131072)

outputs = llm.generate(
    ["用一句话解释什么是 GQA。"],
    sampling_params={"temperature": 1.0, "top_p": 0.95, "max_new_tokens": 1024,
                     "skip_special_tokens": False},
)
print(outputs)

Common pitfalls

GLIBCXX_3.4.31 not found (only if you hit it — not universal): conda Python ships an older libstdc++ than the SGLang wheel was built against. Force-load the system one: LD_PRELOAD=/lib/x86_64-linux-gnu/libstdc++.so.6 python -m sglang.launch_server ...

When NOT to use

Production serving without batch-eval needs → minicpm5-deploy-vllm is simpler
One-shot Python → minicpm5-deploy-transformers
No NVIDIA GPU → minicpm5-deploy-llama-cpp / minicpm5-deploy-mlx

Reference

docs/deployment/sglang.md

minicpm5-deploy-sglang