| <p align="center"> | |
| <img src="https://dscache.tencent-cloud.cn/upload/uploader/hunyuan-64b418fd052c033b228e04bc77bbc4b54fd7f5bc.png" width="400"/> <br> | |
| </p><p></p> | |
| <p align="center"> | |
| 🫣 <a href="https://huggingface.co/tencent/Hunyuan-A13B-Instruct"><b>Hugging Face</b></a> | | |
| 🖥️ <a href="https://llm.hunyuan.tencent.com/" style="color: red;"><b>Official Website</b></a> | | |
| 🕖 <a href="https://cloud.tencent.com/product/hunyuan"><b>HunyuanAPI</b></a> | | |
| 🕹️ <a href="https://hunyuan.tencent.com/?model=hunyuan-a13b"><b>Demo</b></a> | | |
| <img src="https://avatars.githubusercontent.com/u/109945100?s=200&v=4" width="16"/> <a href="https://modelscope.cn/models/Tencent-Hunyuan/Hunyuan-A13B-Instruct"><b>ModelScope</b></a> | |
| </p> | |
| <p align="center"> | |
| <a href="https://github.com/Tencent/Hunyuan-A13B"><b>GITHUB</b></a> | |
| </p> | |
| ## 模型介绍 | |
| 随着人工智能技术的快速发展,大型语言模型(LLMs)在自然语言处理、计算机视觉和科学任务等领域取得了显著进展。然而,随着模型规模的扩大,如何在保持高性能的同时优化资源消耗成为一个关键挑战。为了应对这一挑战,我们研究了混合专家(MoE)模型,当前亮相的 Hunyuan-A13B 模型,拥有800亿总参数和130亿激活参数。不仅在效果上达到了高标准,而且在尺寸上也做到了极致的优化,成功平衡了模型性能与资源占用。 | |
| ### 核心特性与优势 | |
| - **小参数量,高性能**:仅激活130亿参数(总参数量800亿),即可在多样化基准任务中媲美更大规模模型的竞争力表现 | |
| - **混合推理支持**:同时支持快思考和慢思考两种模式,支持用户灵活选择 | |
| - **超长上下文理解**:原生支持256K上下文窗口,在长文本任务中保持稳定性能 | |
| - **增强Agent能力**:优化Agent能力,在BFCL-v3、τ-Bench等智能体基准测试中领先 | |
| - **高效推理**:采用分组查询注意力(GQA)策略,支持多量化格式,实现高效推理 | |
| ### 为何选择Hunyuan-A13B? | |
| 作为兼具强大性能与计算效率的大模型,Hunyuan-A13B是研究者与开发者在资源受限条件下追求高性能的理想选择。无论学术研究、高性价比AI解决方案开发,还是创新应用探索,本模型都能提供强大的基础支持。 | |
| | |
| ## 新闻 | |
| <br> | |
| * 2025.6.26 我们在Hugging Face开源了 **Hunyuan-A13B-Instruct**,**Hunyuan-A13B-Pretrain**, **Hunyuan-A13B-Instruct-FP8**, **Hunyuan-A13B-Instruct-GPTQ-Int4**。并发布了技术报告和训练推理操作手册,详细介绍了模型能力和训练与推理的操作。 | |
| ## 模型结构 | |
| Hunyuan-A13B采用了细粒度混合专家(Fine-grained Mixture of Experts,Fine-grained MoE)架构,包含800亿参数和130亿激活参数,累计训练了超过 20T tokens。该模型支持 256K 的上下文长度,以下为模型结构细节: | |
| * 总参数: 80B | |
| * 激活参数: 13B | |
| * 层数: 32 | |
| * Attention Heads: 32 | |
| * 共享专家数: 1 | |
| * 非共享专家数: 64 | |
| * 路由策略: Top-8 | |
| * 激活函数: SwiGLU | |
| * 隐层维度: 4096 | |
| * 专家隐层维度: 3072 | |
| ## Benchmark评估榜单 | |
| **Hunyuan-A13B-Pretrain** 在 12/14 个任务上超越了Hunyuan上一代52B激活参数的MoE模型Hunyuan-Large,证实了它在预训练任务上出色的能力。与业界更大参数量的Dense和MoE模型相比, Hunyuan-A13B在多个代码和数学任务上都取得了最高分数。在MMLU, MMLU-PRO等诸多众聚合任务上, Hunyuan-A13B达到了与Qwen3-A22B模型同等的水平,表现出优秀的综合能力。 | |
| | Model | Hunyuan-Large | Qwen2.5-72B | Qwen3-A22B | Hunyuan-A13B | | |
| |------------------|---------------|--------------|-------------|---------------| | |
| | MMLU | 88.40 | 86.10 | 87.81 | 88.17 | | |
| | MMLU-Pro | 60.20 | 58.10 | 68.18 | 67.23 | | |
| | MMLU-Redux | 87.47 | 83.90 | 87.40 | 87.67 | | |
| | BBH | 86.30 | 85.80 | 88.87 | 87.56 | | |
| | SuperGPQA | 38.90 | 36.20 | 44.06 | 41.32 | | |
| | EvalPlus | 75.69 | 65.93 | 77.60 | 78.64 | | |
| | MultiPL-E | 59.13 | 60.50 | 65.94 | 69.33 | | |
| | MBPP | 72.60 | 76.00 | 81.40 | 83.86 | | |
| | CRUX-I | 57.00 | 57.63 | - | 70.13 | | |
| | CRUX-O | 60.63 | 66.20 | 79.00 | 77.00 | | |
| | MATH | 69.80 | 62.12 | 71.84 | 72.35 | | |
| | CMATH | 91.30 | 84.80 | - | 91.17 | | |
| | GSM8k | 92.80 | 91.50 | 94.39 | 91.83 | | |
| | GPQA | 25.18 | 45.90 | 47.47 | 49.12 | | |
| **Hunyuan-A13B-Instruct** 在多项基准测试中取得了极具有竞争力的表现,尤其是在数学、科学、agent等领域。我们与一些强力模型进行了对比,结果如下所示。 | |
| | Topic | Bench | OpenAI-o1-1217 | DeepSeek R1 | Qwen3-A22B | Hunyuan-A13B-Instruct | | |
| |:-------------------:|:-----------------------------:|:-------------:|:------------:|:-----------:|:---------------------:| | |
| | **Mathematics** | AIME 2024<br>AIME 2025<br>MATH | 74.3<br>79.2<br>96.4 | 79.8<br>70<br>94.9 | 85.7<br>81.5<br>94.0 | 87.3<br>76.8<br>94.3 | | |
| | **Science** | GPQA-Diamond<br>OlympiadBench | 78<br>83.1 | 71.5<br>82.4 | 71.1<br>85.7 | 71.2<br>82.7 | | |
| | **Coding** | Livecodebench<br>Fullstackbench<br>ArtifactsBench | 63.9<br>64.6<br>38.6 | 65.9<br>71.6<br>44.6 | 70.7<br>65.6<br>44.6 | 63.9<br>67.8<br>43 | | |
| | **Reasoning** | BBH<br>DROP<br>ZebraLogic | 80.4<br>90.2<br>81 | 83.7<br>92.2<br>78.7 | 88.9<br>90.3<br>80.3 | 89.1<br>91.1<br>84.7 | | |
| | **Instruction<br>Following** | IF-Eval<br>SysBench | 91.8<br>82.5 | 88.3<br>77.7 | 83.4<br>74.2 | 84.7<br>76.1 | | |
| | **Text<br>Creation**| LengthCtrl<br>InsCtrl | 60.1<br>74.8 | 55.9<br>69 | 53.3<br>73.7 | 55.4<br>71.9 | | |
| | **NLU** | ComplexNLU<br>Word-Task | 64.7<br>67.1 | 64.5<br>76.3 | 59.8<br>56.4 | 61.2<br>62.9 | | |
| | **Agent** | BDCL v3<br> τ-Bench<br>ComplexFuncBench<br> $C^3$-Bench | 67.8<br>60.4<br>47.6<br>58.8 | 56.9<br>43.8<br>41.1<br>55.3 | 70.8<br>44.6<br>40.6<br>51.7 | 78.3<br>54.7<br>61.2<br>63.5 | | |
| ## 推理和部署 | |
| HunyuanLLM可以采用vLLM,sglang或TensorRT-LLM部署。为了简化部署过程HunyuanLLM提供了预构建docker镜像。 | |
| ## 使用TensorRT-LLM推理 | |
| ### BF16部署 | |
| #### Step1:执行推理 | |
| #### 方式1:命令行推理 | |
| 下面我们展示一个代码片段,采用`TensorRT-LLM`快速请求chat model: | |
| 修改 examples/pytorch/quickstart_advanced.py 中如下代码: | |
| ```python | |
| from tensorrt_llm import SamplingParams | |
| from tensorrt_llm._torch import LLM | |
| from tensorrt_llm._torch.pyexecutor.config import PyTorchConfig | |
| from tensorrt_llm.llmapi import (EagleDecodingConfig, KvCacheConfig, | |
| MTPDecodingConfig) | |
| prompt = "Write a short summary of the benefits of regular exercise" | |
| def main(): | |
| args = parse_arguments() | |
| llm, sampling_params = setup_llm(args) | |
| new_prompts = [] | |
| if args.apply_chat_template: | |
| messages = [{"role": "user", "content": f"{prompt}"}] | |
| new_prompts.append(llm.tokenizer.apply_chat_template( | |
| messages, tokenize=False, add_generation_prompt=True) | |
| ) | |
| outputs = llm.generate(new_prompts, sampling_params) | |
| for i, output in enumerate(outputs): | |
| prompt = output.prompt | |
| generated_text = output.outputs[0].text | |
| print(f"[{i}] Prompt: {prompt!r}, Generated text: {generated_text!r}") | |
| ``` | |
| 运行方式: | |
| ```shell | |
| python3 quickstart_advanced.py --model_dir "HunyuanLLM模型路径" --tp_size 4 --apply_chat_template | |
| ``` | |
| #### 方式2:服务化推理 | |
| 下面我们展示使用`TensorRT-LLM`服务化的方式部署模型和请求。 | |
| ```shell | |
| model_path="HunyuanLLM模型路径" | |
| trtllm-serve <model_path> [--backend pytorch --tp_size <tp> --ep_size <ep> --host <host> --port <port>] | |
| ``` | |
| 服务启动成功后, 运行请求脚本: | |
| ```python | |
| ### OpenAI Chat Client | |
| from openai import OpenAI | |
| client = OpenAI( | |
| base_url="http://localhost:8000/v1", | |
| api_key="tensorrt_llm", | |
| ) | |
| response = client.chat.completions.create( | |
| model="default", | |
| messages=[{ | |
| "role": "user", | |
| "content": "Write a short summary of the benefits of regular exercise" | |
| }], | |
| max_tokens=4096, | |
| ) | |
| print(response) | |
| ``` | |
| #### FP8/Int4量化模型部署: | |
| 目前 TensorRT-LLM 的 fp8 和 int4 量化模型正在支持中,敬请期待。 | |
| ## 使用vLLM推理 | |
| ### Docker: | |
| 为了简化部署过程,HunyuanLLM提供了预构建docker镜像: | |
| [hunyuaninfer/hunyuan-large:hunyuan-moe-A13B-vllm](https://hub.docker.com/r/hunyuaninfer/hunyuan-large/tags) 。您只需要下载模型文件并用下面代码启动docker即可开始推理模型。 | |
| ```shell | |
| # 拉取 | |
| docker pull hunyuaninfer/hunyuan-large:hunyuan-moe-A13B-vllm | |
| # 起镜像 | |
| docker run --name hunyuanLLM_infer -itd --privileged --user root --net=host --ipc=host --gpus=8 hunyuaninfer/hunyuan-large:hunyuan-moe-A13B-vllm | |
| ``` | |
| 注: Docker容器权限管理。以上代码采用特权模式(--privileged)启动Docker容器会赋予容器较高的权限,增加数据泄露和集群安全风险。建议在非必要情况下避免使用特权模式,以降低安全威胁。对于必须使用特权模式的场景,应进行严格的安全评估,并实施相应的安全监控、加固措施。 | |
| ### BF16部署 | |
| BF16可以在2张显存超过80G的GPU卡上部署,如果长文推荐TP4。按如下步骤执行: | |
| 运行命令前请先设置如下环境变量: | |
| ```shell | |
| export MODEL_PATH=PATH_TO_MODEL | |
| ``` | |
| #### Step1:执行推理 | |
| #### 方式1:命令行推理 | |
| 下面我们展示一个代码片段,采用`vLLM`快速请求chat model: | |
| 注: vLLM组件远程代码执行防护。下列代码中vLLM组件的trust-remote-code配置项若被启用,将允许加载并执行来自远程模型仓库的代码,这可能导致恶意代码的执行。除非业务需求明确要求,否则建议该配置项处于禁用状态,以降低潜在的安全威胁。 | |
| ```python | |
| import os | |
| from typing import List, Optional | |
| from vllm import LLM, SamplingParams | |
| from vllm.inputs import PromptType | |
| from transformers import AutoTokenizer | |
| model_path=os.environ.get('MODEL_PATH') | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) | |
| llm = LLM(model=model_path, | |
| tokenizer=model_path, | |
| trust_remote_code=True, | |
| dtype='bfloat16', | |
| tensor_parallel_size=4, | |
| gpu_memory_utilization=0.9) | |
| sampling_params = SamplingParams( | |
| temperature=0.7, top_p=0.8, max_tokens=4096, top_k=20, repetition_penalty=1.05) | |
| messages = [ | |
| { | |
| "role": "system", | |
| "content": "You are a helpful assistant.", | |
| }, | |
| {"role": "user", "content": "Write a short summary of the benefits of regular exercise"}, | |
| ] | |
| tokenized_chat = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt") | |
| dummy_inputs: List[PromptType] = [{ | |
| "prompt_token_ids": batch | |
| } for batch in tokenized_chat.numpy().tolist()] | |
| outputs = llm.generate(dummy_inputs, sampling_params) | |
| # Print the outputs. | |
| for output in outputs: | |
| prompt = output.prompt | |
| generated_text = output.outputs[0].text | |
| print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") | |
| ``` | |
| #### 方式2:服务化推理 | |
| 下面我们展示使用`vLLM`服务化的方式部署模型并请求 | |
| 在主节点上运行: | |
| ```shell | |
| export VLLM_HOST_IP=${LOCAL_IP} | |
| ``` | |
| 接着我们启动服务,运行 : | |
| ```shell | |
| cd inference | |
| sh run_server.sh | |
| ``` | |
| 运行`run_server.sh`成功后, 运行请求脚本: | |
| ```shell | |
| sh openapi.sh | |
| ``` | |
| 注意修改`openapi.sh`中的`${LOCAL_IP}`和`${MODEL_PATH}`为服务对应值。 | |
| ### 量化模型部署: | |
| 本部分介绍采用vLLM部署量化后模型的流程。 | |
| 镜像:部署镜像同BF16。 | |
| #### Int8量化模型部署: | |
| 部署Int8-weight-only版本HunYuan-A13B模型只需设置`run_server_int8.sh`中的环境变量: | |
| ```SHELL | |
| export MODEL_PATH=PATH_TO_BF16_MODEL | |
| ``` | |
| 接着我们启动Int8服务。运行: | |
| ```shell | |
| sh run_server_int8.sh | |
| ``` | |
| 运行`run_server_int8.sh`成功后, 运行请求脚本: | |
| ```shell | |
| sh openapi.sh | |
| ``` | |
| #### Int4量化模型部署: | |
| 部署Int4-weight-only版本HunYuan-A13B模型只需设置`run_server_int4.sh`中的环境变量,采用GPTQ方式: | |
| ```SHELL | |
| export MODEL_PATH=PATH_TO_INT4_MODEL | |
| ``` | |
| 接着我们启动Int4服务。运行: | |
| ```shell | |
| sh run_server_int4.sh | |
| ``` | |
| 运行`run_server_int4.sh`成功后, 运行请求脚本: | |
| ```shell | |
| sh openapi.sh | |
| ``` | |
| #### FP8量化模型部署: | |
| 部署W8A8C8版本HunYuan-A13B模型只需设置`run_server_int8.sh`中的环境变量: | |
| ```shell | |
| export MODEL_PATH=PATH_TO_FP8_MODEL | |
| ``` | |
| 接着我们启动FP8服务。运行: | |
| ```shell | |
| sh run_server_fp8.sh | |
| ``` | |
| 运行`run_server_fp8.sh`成功后, 运行请求脚本: | |
| ```shell | |
| sh openapi.sh | |
| ``` | |
| ### 性能评估: | |
| 本部分介绍采用vLLM部署各个模型(原始模型和量化模型)的效率测试结果,包括不同Batchsize下的推理速度(tokens/s), 测试环境(腾讯云,H80(96G)GPU x 卡数): | |
| 测试命令: | |
| ```python | |
| python3 benchmark_throughput.py --backend vllm \ | |
| --input-len 2048 \ | |
| --output-len 14336 \ | |
| --model $MODEL_PATH \ | |
| --tensor-parallel-size $TP \ | |
| --use-v2-block-manager \ | |
| --async-engine \ | |
| --trust-remote-code \ | |
| --num_prompts $BATCH_SIZE \ | |
| --max-num-seqs $BATCH_SIZE | |
| ``` | |
| | 推理框架 | 模型 | 部署卡数 | input_length | batch=1 | batch=16 | batch=32 | | |
| |------|-----------------------------|-----------|-------------------------|---------------------|----------------------|----------------------| | |
| | vLLM | Hunyuan-A13B-Instruct | 8 | 2048 | 190.84 | 1246.54 | 1981.99 | | |
| | vLLM | Hunyuan-A13B-Instruct | 4 | 2048 | 158.90 | 779.10 | 1301.75 | | |
| | vLLM | Hunyuan-A13B-Instruct | 2 | 2048 | 111.72 | 327.31 | 346.54 | | |
| | vLLM | Hunyuan-A13B-Instruct(int8 weight only) | 2 | 2048 | 109.10 | 444.17 | 721.93 | | |
| | vLLM | Hunyuan-A13B-Instruct(W8A8C8-FP8) | 2 | 2048 | 91.83 | 372.01 | 617.70 | | |
| | vLLM | Hunyuan-A13B-Instruct(W8A8C8-FP8) | 1 | 2048 | 60.07 | 148.80 | 160.41 | | |
| ## 使用sglang推理 | |
| ### BF16部署 | |
| #### Step1:执行推理 | |
| #### 方式1:命令行推理 | |
| 下面我们展示一个代码片段,采用`sglang`快速请求chat model: | |
| ```python | |
| import sglang as sgl | |
| from transformers import AutoTokenizer | |
| model_path=os.environ.get('MODEL_PATH') | |
| tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True) | |
| messages = [ | |
| { | |
| "role": "system", | |
| "content": "You are a helpful assistant.", | |
| }, | |
| {"role": "user", "content": "Write a short summary of the benefits of regular exercise"}, | |
| ] | |
| prompts = [] | |
| prompts.append(tokenizer.apply_chat_template( | |
| messages, | |
| tokenize=False, | |
| add_generation_prompt=True | |
| )) | |
| print(prompts) | |
| llm = sgl.Engine( | |
| model_path=model_path, | |
| tp_size=4, | |
| trust_remote_code=True, | |
| mem_fraction_static=0.7, | |
| ) | |
| sampling_params = {"temperature": 0.7, "top_p": 0.8, "top_k": 20, "max_new_tokens": 4096} | |
| outputs = llm.generate(prompts, sampling_params) | |
| for prompt, output in zip(prompts, outputs): | |
| print(f"Prompt: {prompt}\nGenerated text: {output['text']}") | |
| ``` | |
| #### 方式2:服务化推理 | |
| 下面我们展示使用`sglang`服务化的方式部署模型和请求。 | |
| ```shell | |
| model_path="HunyuanLLM模型路径" | |
| python3 -u -m sglang.launch_server \ | |
| --model-path $model_path \ | |
| --tp 4 \ | |
| --trust-remote-code \ | |
| ``` | |
| 服务启动成功后, 运行请求脚本: | |
| ```python | |
| import openai | |
| client = openai.Client( | |
| base_url="http://localhost:30000/v1", api_key="EMPTY") | |
| response = client.chat.completions.create( | |
| model="default", | |
| messages= [ | |
| {"role": "user", "content": "Write a short summary of the benefits of regular exercise"}, | |
| ], | |
| temperature=0.7, | |
| max_tokens=4096, | |
| extra_body={"top_p": 0.8, "top_k": 20} | |
| ) | |
| print(response) | |
| ``` | |
| #### FP8/Int4量化模型部署: | |
| 目前 sglang 的 fp8 和 int4 量化模型正在支持中,敬请期待。 | |
| ## 交互式Demo Web | |
| hunyuan-A13B 现已开放网页demo。访问 https://hunyuan.tencent.com/?model=hunyuan-a13b 即可简单体验我们的模型。 | |
| <br> | |
| ## 引用 | |
| 如果你觉得我们的工作对你有帮助,欢迎引用我们的<a href="report/Hunyuan_A13B_Technical_Report.pdf">技术报告</a>! | |
| <br> | |
| ## 联系我们 | |
| 如果你想给我们的研发和产品团队留言,欢迎联系我们腾讯混元LLM团队。你可以通过邮件(hunyuan[email protected])联系我们。 |