Higher performance OpenAI LLM service than vLLM serve: A pure C++ high-performance OpenAI LLM service implemented with GPRS+TensorRT-LLM+Tokenizers.cpp, supporting chat and function call, AI agents, distributed multi-GPU inference, multimodal capabilities, and a Gradio chat interface.
# Add to your Claude Code skills
git clone https://github.com/NetEase-Media/grps_trtllmLast scanned: 5/30/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-05-30T16:00:09.964Z",
"npmAuditRan": true,
"pipAuditRan": true
}grps_trtllm is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by NetEase-Media. Higher performance OpenAI LLM service than vLLM serve: A pure C++ high-performance OpenAI LLM service implemented with GPRS+TensorRT-LLM+Tokenizers.cpp, supporting chat and function call, AI agents, distributed multi-GPU inference, multimodal capabilities, and a Gradio chat interface. It has 159 GitHub stars.
Yes. grps_trtllm passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/NetEase-Media/grps_trtllm" and add it to your Claude Code skills directory (see the Installation section above).
grps_trtllm is primarily written in Python. It is open-source under NetEase-Media on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh grps_trtllm against similar tools.
No comments yet. Be the first to share your thoughts!
GRPS + TensorRT-LLM
实现纯C++版,相比vllm serve更优性能的OpenAI LLM服务,支持Chat、Ai-agent、Multi-modal
、多卡推理等。
grps接入trtllm
实现更高性能的、支持OpenAI模式访问、支持Ai-agent以及多模态的LLM
服务:
C++实现完整LLM服务,包含tokenizer(支持huggingface, sentencepiecetokenizer)、llm推理
、vit等部分。grps的自定义http功能实现OpenAI接口协议,支持chat和function call模式。LLM的prompt构建风格以及生成结果的解析风格,以实现不同LLM的chat
和function call模式,支持llama-indexai-agent。tensorrt推理后端与opencv库,支持多模态LLM。inflight batching、multi-gpu、paged attention、kv-cache reuse、lookahead decoding等
TensorRT-LLM推理加速技术。triton_server <--> tokenizer_backend <--> trtllm_backend之间的进程间通信,纯C++实现,性能有稳定的提升。欢迎各位使用和提issue ,欢迎提交pr支持新的模型,感谢star⭐️。
2025-05-14
2025-05-07
2025-04-17
2025-03-25
2025-03-22
2025-03-20
2025-03-06
2025-03-04
2025-02-28
2025-02-24
2025-02-23
2025-02-21
2025-02-05
2025-01-24
2025-01-08
2024-12-24
2024-12-19
2024-12-17
支持的文本LLM:
| supported model | llm_styler | chat | function_call | doc |
|---|---|---|---|---|
| Qwen3 | qwen3 | ✅ | ✅ | qwen3 |
| DeepSeek-R1-DistillTinyR1-32B-Preview | deepseek-r1 | ✅ | ❌ | deepseek-r1-distill |
| QwQ-32BQwQ-32B-AWQ | qwq | ✅ | ✅ | qwq |
| QwQ-32B-Preview | qwq-preview | ✅ | ❌ | qwq-preview |
| Qwen2.5-1MQwen2.5-CoderQwen2.5-MathQwen2.5 | qwen2.5 | ✅ | ✅ | qwen2.5 |
| Qwen1.5-ChatQwen1.5-Moe-ChatQwen2-InstructQwen2-Moe-Instruct | qwen | ✅ | ✅ | qwen2 |
| chatglm3 | chatglm3 | ✅ | ✅ | chatglm3 |
| glm4 | glm4 | ✅ | ✅ | glm4 |
| internlm2_5-chatinternlm2-chat | internlm2 | ✅ | ✅ | internlm2.5 |
| llama-3-instructllama-3.1-instruct | llama3 | ✅ | ❌ | llama3 |
| phi-4 | phi4 | ✅ | ❌ | phi4 |
| Phi-3, Phi-3.5 | phi3 | ✅ | ❌ | phi3 |
| gemma-3(experimental) | gemma3 | ✅ | ❌ | gemma-3 |
支持的多模态LLM(少部分模型vit无法通过纯c++实现):
| supported model | llm_styler | vit | vit_type | chat | function_call | doc |
|---|---|---|---|---|---|---|
| InternVL3 | internvl3 | internvl2 | c++ | ✅ | ❌ | internvl3 |
| MiniCPM-V-2_6 | minicpmv | minicpmv | py | ✅ | ❌ | minicpmv |
| Janus-Pro | janus-pro | janus-pro | c++ | ✅ | ❌ | janus-pro |
| InternVideo2.5 | intern-video2.5 | intern-video2.5 | py | ✅ | ❌ | intern-video2.5 |
| InternVL2_5InternVL2_5-MPO | internvl2.5 | internvl2 | c++ | ✅ | ❌ | internvl2.5 |
| InternVL2-2BInternVL2-8BInternVL2-26B | internvl2-internlm2 | internvl2 | c++ | ✅ | ❌ | internvl2 |
| InternVL2-1B | internvl2-qwen2 | internvl2 | c++ | ✅ | ❌ | internvl2 |
| InternVL2-4B | internvl2-phi3 | internvl2 | c++ | ✅ | ❌ | internvl2 |
| olmOCR | qwen2vl | qwen2vl | c++ | ✅ | ❌ | olm-ocr |
| Qwen2-VL-Instruct | qwen2vl | qwen2vl | c++ | ✅ | ❌ | qwen2vl |
| Qwen-VL-ChatQwen-VL | qwenvl | qwenvl | c++ | ✅ | ❌ | qwenvl |
|-- client # 客户端样例
|-- conf # 配置文件
| |-- inference*.yml # 各类llm推理配置
| |-- server.yml # 服务配置
|-- data # 数据文件
|-- docker # docker镜像构建
|-- docs # 文档
|-- processors # 远程处理器
|-- second_party # grps框架依赖
|-- src # 自定义源码
| |-- tensorrt # tensorrt推理后端
| |-- vit # vit实现
| |-- constants.cc/.h # 常量定义
| |-- customized_inferer.cc/.h # 自定义推理器
| |-- llm_styler.cc/.h # LLM风格定义,prompt构建,结果解析
| |-- tokenizer.cc/.h # Tokenizer实现
| |-- trtllm_model_instance.cc/.h # TensorRT-LLM模型实例
| |-- trtllm_model_state.cc/.h # TensorRT-LLM模型状态
| |-- utils.cc/.h # 工具
| |-- main.cc # 本地单元测试
|-- third_party # 第三方依赖
|-- tools # 工具
|-- build.sh # 构建脚本
|-- CMakelists.txt # 工程构建文件
|-- .clang-format # 代码格式化配置文件
|-- .config # 工程配置文件,包含一些工程配置开关
以qwen2.5-instruct为例。更多llm示例见模型列表,拉取代码与创建容器步骤相同。
git clone https://github.com/NetEase-Media/grps_trtllm.git
cd grps_trtllm
git submodule update --init --recursive
使用```registry.cn-hangzhou.aliyuncs.com/opengrps/grps_gpu: