--- title: Qwen2.5-7B-Instruct-Int8-W8A16 canonical_url: "https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-Int8-W8A16" md_url: "https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-Int8-W8A16.md" repository: okwinds/Qwen2.5-7B-Instruct-Int8-W8A16 chinese_name: "通义千问Qwen2.5-7B-Instruct-Int8-W8A16量化高精校准" last_updated: 2025-04-17 license: qwen pipeline_tag: text-generation tasks: - text-generation model_type: - qwen2 architectures: - Qwen2ForCausalLM base_model: - Qwen/Qwen2.5-7B-Instruct base_model_relation: finetune parameters: 2.7B tensor_type: - I32 - BF16 - I64 library_name: - safetensors - pytorch frameworks: - pytorch language: - en - zh inference_backends: - "deploy_task text/emb" - "lmdeploy 0.9.1" - "lmdeploy_turbomind 0.9.1" - "sglang 0.5.2" - "vllm 0.9.2" downloads: 2329 stars: 7 tags: - chat --- # Qwen2.5-7B-Instruct-Int8-W8A16 > Qwen2.5-7B-Instruct-Int8-W8A16 - okwinds 在 ModelScope 开源的模型。通义千问Qwen2.5-7B-Instruct-Int8-W8A16量化高精校准 okwinds/Qwen2.5-7B-Instruct-Int8-W8A16 是 ModelScope 魔搭社区上的 2.7B 参数text-generation模型，采用 qwen 许可，基于 Qwen/Qwen2.5-7B-Instruct 构建，可用 deploy_task text/emb、lmdeploy 0.9.1、lmdeploy_turbomind 0.9.1 部署。 - **Repository**: okwinds/Qwen2.5-7B-Instruct-Int8-W8A16 - **License**: qwen - **Tasks**: text-generation - **Parameters**: 2.7B - **Base model**: Qwen/Qwen2.5-7B-Instruct - **Inference backends**: deploy_task text/emb, lmdeploy 0.9.1, lmdeploy_turbomind 0.9.1, sglang 0.5.2, vllm 0.9.2 - **Tags**: chat - **Downloads**: 2329 - **Stars**: 7 - **Last updated**: 2025-04-17 Source: https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-Int8-W8A16 --- # 通义千问Qwen2.5-7B-Instruct-Int8-W8A16量化高精校准 原模型 [qwen/Qwen2.5-7B-Instruct](https://www.modelscope.cn/models/qwen/Qwen2.5-7B-Instruct) 友情提醒，推理本模型时，请按照如下官方指引： > 注意：建议在运行这些模型时设置适当的温度（0.5 到 0.7 之间），否则您可能会遇到无休止的重复或不连贯的输出问题。 📖 关于量化损失方面的研究，可阅读公众号“觉察流”文章👇
### 《[Reason Model 的“瘦身计划”：量化技术的得与失](https://mp.weixin.qq.com/s/NMGq4UUkfo8GMix5LHnWCg)》 ```python # download model from modelscope import snapshot_download model_dir = snapshot_download('okwinds/Qwen2.5-7B-Instruct-Int8-W8A16') ``` ## Qwen2.5 同系列量化模型 | **Model** | GPTQ Compression & FP8 | GGUF | **MLX** (macOS) | | ------------ |----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| -------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | *3B params* | - | [3B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-3B-Instruct-GGUF-V3-LOT) | [3B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-3B-Instruct-MLX-4bit)
[3B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-3B-Instruct-MLX-8bit) | | *7B params* | [7B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-Int4-W4A16)
[7B-Int8-W8A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-Int8-W8A16)
[7B-FP8-A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-FP8-A16)
[Coder-7B-Instruct-FP8](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-FP8) | [7B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-GGUF-V3-LOT) | [7B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-MLX-4bit)
[7B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-7B-Instruct-MLX-8bit) | | *Coder 7B params* | [Coder-7B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-Int4-W4A16)
[Coder-7B-Int8-W8A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-Int8-W8A16)
[Coder-7B-FP8-A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-FP8-A16) | [Coder-7B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-GGUF-V3-LOT) | [Coder-7B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-MLX-4bit)
[Coder-7B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-7B-Instruct-MLX-8bit) | | *14B params* | [14B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-Int4-W4A16)
[14B-Int8-W8A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-Int8-W8A16)
[14B-FP8-A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-FP8-A16) | [14B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-GGUF-V3-LOT) | [14B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-MLX-4bit)
[14B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-MLX-8bit) | | *Coder 14B params* | [Coder-14B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-14B-Instruct-Int4-W4A16)
[Coder-14B-Int8-W8A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-14B-Instruct-Int8-W8A16) | [Coder-14B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-14B-Instruct-GGUF-V3-LOT) | [Coder-14B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-14B-Instruct-MLX-4bit)
[Coder-14B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-14B-Instruct-MLX-8bit) | | *32B params* | [32B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-Int4-W4A16)
[32B-Int8-W8A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-Int8-W8A16)
[32B-FP8-A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-FP8-A16) | [32B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-GGUF-V3-LOT) | [32B-MLX-2bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-MLX-2bit)
[32B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-MLX-4bit)
[32B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-32B-Instruct-MLX-8bit) | | *Coder 32B params* | [Coder-32B-Int4-W4A16](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-32B-Instruct-Int4-W4A16)
[Coder-32B-FP8](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-32B-Instruct-FP8) | [Coder-32B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-32B-Instruct-GGUF-V3-LOT) | [Coder-32B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-32B-Instruct-MLX-4bit)
[Coder-32B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-Coder-32B-Instruct-MLX-8bit) | | *QwQ-Preview 32B params* | [QwQ-32B-Preview-Int4-W4A16](https://www.modelscope.cn/models/okwinds/QwQ-32B-Preview-Int4-W4A16)
[QwQ-32B-Preview-FP8](https://www.modelscope.cn/models/okwinds/QwQ-32B-Preview-FP8) | [QwQ-32B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/QwQ-32B-Preview-GGUF-V3-LOT) | [QwQ-32B-MLX-4bit](https://www.modelscope.cn/models/okwinds/QwQ-32B-Preview-MLX-4bit)
[QwQ-32B-MLX-8bit](https://www.modelscope.cn/models/okwinds/QwQ-32B-Preview-MLX-8bit) | | *72B params* | - | [72B-GGUF-V3-LOT](https://www.modelscope.cn/models/okwinds/Qwen2.5-72B-Instruct-GGUF-V3-LOT) | [72B-MLX-2bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-72B-Instruct-MLX-2bit)
[72B-MLX-4bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-72B-Instruct-MLX-4bit)
[72B-MLX-8bit](https://www.modelscope.cn/models/okwinds/Qwen2.5-72B-Instruct-MLX-8bit) | ## 模型概述 > 本模型已升级，请使用精度更高的最新版本，原版本归档至 tag v1.0.0。
更多精度详情，请关注 Benchmark 部分数据。 通义千问2.5-7B-Instruct-INT8-W8A16 是一个基于 通义千问2.5-7B-Instruct 的 INT8 量化并校准的模型。 - **模型名称:** Qwen2.5-7B-Instruct-Int8-W8A16 - **模型架构:** Qwen2.5 - **权重量化:** INT8 该模型通过将 [Qwen2.5-7B-Instruct](https://www.modelscope.cn/models/qwen/Qwen2.5-7B-Instruct) 的权重量化为 INT8 数据类型而实现。 量化过程将每个参数从 16bit 减少到 8bit，将模型占用磁盘空间大小，以及推理时加载模型需要的GPU显存空间，减少到了大约为原模型的1/2。 量化过程中，只有 transformer 中的 Linear 层权重是量化的，其他层均保持为原模型数据类型，采用混合精度的计算方式，尽可能减少模型量化后所带来的精度损失。 在量化过程中，使用了[AutoGPTQ](https://github.com/AutoGPTQ/AutoGPTQ)，采用 Symmetric group-wise 方式量化。并在量化过程中，进行了数据校准，以提升模型生成精度，修复 gptq-int8 模型出现乱吐字和“!”无限循环的问题。 > 注意：本模型需要 compute capability > 8.0（Ampere、Ada Lovelace、Hopper 架构）的 Nvidia GPU 来支持 INT8 **混合精度**计算。 ## 部署推理 #### 推荐使用 vLLM Openai api 兼容模式 ```bash >>> python -m vllm.entrypoints.openai.api_server --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.9 --served-model-name "Qwen2.5-7B-Instruct-Int8-W8A16" --model "/home/gavin/llm/Qwen2.5-7B-Instruct-Int8-W8A16" ``` ## Benchmark - INT8(W8A16) 的量化方法，对比原模型在精度上的表现，几乎无损失。在做数据校准的时候，也最大限度的保有了BF16原模型的精度。 - 本模型的最新版本对精度进行了更广泛的校准。尤其对推理，CoT，指令遵循，Agent等能力做了强化提升。 | Benchmark | Qwen2.5-7B-Instruct | Int8-W8A16-v1.0.0 | Recovery-v1.0.0 | _Int8-W8A16-master(New)_ | Recovery-master(New) | | -------------------- | ------------------- | ----------------- | --------------- | ------------------- |---------------------------------------------------------------------| | **Overall** | 70.69 | 70.94 | 100.36% | 71.09 | 100.57% | | Exam | 77.90 | 77.97 | 100.09% | 77.82 | 99.90% | | Language | 53.97 | 53.62 | 99.35% | 54.33 | 100.66% | | Knowledge | 47.07 | 48.08 | 102.15% | 47.69 | 101.32% | | Understanding | 89.74 | 89.78 | 100.04% | 89.79 | 100.06% | | Coding | 79.32 | 80.02 | 100.88% | 80.63 | 101.65% | | Reasoning | 69.22 | 69.00 | 99.69% | 69.21 | 99.99% | | Instruct & Tools | 77.59 | 78.12 | 100.68% | 78.15 | 100.72% | | **Exam** | - | - | - | - | - | | C-EVAL | 78.52 | 78.37 | 99.81% | 78.21 | 99.61% | | MMLU | 74.19 | 74.22 | 100.04% | 74.15 | 99.95% | | CMMLU | 78.78 | 78.92 | 100.18% | 78.82 | 100.05% | | GaokaoBench | 80.10 | 80.36 | 100.32% | 80.11 | 100.01% | | **Language** | - | - | - | - | - | | WiC | 55.80 | 55.49 | 99.44% | 55.33 | 99.16% | | WSC | 70.19 | 69.23 | 98.63% | 72.12 | 102.75% | | Afqmc-dev | 70.64 | 70.57 | 99.90% | 70.53 | 99.84% | | Tydiqa Goldp | 19.26 | 19.20 | 99.69% | 19.32 | 100.31% | | **Knowledge** | - | - | - | - | - | | BoolQ | 84.92 | 85.11 | 100.22% | 84.86 | 99.93% | | TriviaQA | 53.42 | 53.63 | 100.39% | 53.27 | 99.72% | | GPQA Diamond | 31.82 | 35.35 | 111.09% | 34.34 | 107.92% | | Natual Questions | 18.12 | 18.23 | 100.61% | 18.28 | 100.88% | | **Understanding** | - | - | - | - | - | | C3 | 91.51 | 91.78 | 100.30% | 91.84 | 100.36% | | Race-middle | 91.09 | 90.88 | 99.77% | 90.95 | 99.85% | | Race-high | 86.62 | 86.68 | 100.07% | 86.59 | 99.97% | | **Coding** | - | - | - | - | - | | Human Eval | 83.54 | 84.15 | 100.73% | 85.37 | 102.19% | | Sanitized MBPP | 75.10 | 75.88 | 101.04% | 75.88 | 101.04% | | **Reasoning** | - | - | - | - | - | | COPA | 96.00 | 96.00 | 100.00% | 96.00 | 100.00% | | MATH | 54.12 | 54.06 | 99.89% | 54.04 | 99.85% | | GSM8K | 87.26 | 87.11 | 99.83% | 87.64 | 100.44% | | Winogrande | 72.14 | 72.30 | 100.22% | 72.06 | 99.89% | | Hellaswag | 85.38 | 85.46 | 100.09% | 85.49 | 100.13% | | BBH | 70.12 | 69.97 | 99.79% | 70.52 | 100.57% | | TheoremQA | 19.50 | 18.12 | 92.92% | 18.75 | 96.15% | | **Instruct & Tools** | - | - | - | - | - | | IFEval | 78.66 | 79.74 | 101.37% | 80.10 | 101.83% | | TEval | 76.52 | 76.49 | 99.96% | 76.20 | 99.58% | _以上 benchmark 以及量化工作均在 Nvidia 4090 RTX 24GB 显卡上完成。_ ***并发说明*** 在以上 benchmark 基准测试中，采用 vllm 推理，观察到以下基准测试中的并发表现。 - vLLM 0.5.5 - 原模型 BF16 峰值并发，在无 pending 的前提下，达到 running 90 左右。 - 量化后的本模型 INT8(W8A16) 峰值并发，在无 pending 的前提下，达到 running 140 左右。 - vLLM 0.6.1.post2 - 原模型 BF16 峰值并发，在无 pending 的前提下，达到 running 130 左右。 - 量化后的本模型 INT8(W8A16) 峰值并发，在无 pending 的前提下，达到 running 250 左右。
_______________________________ #### _作者在此 👇🏻 扫一扫_