--- title: Qwen2.5-14B-Instruct-Int8-W8A16 canonical_url: "https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-Int8-W8A16" md_url: "https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-Int8-W8A16.md" repository: okwinds/Qwen2.5-14B-Instruct-Int8-W8A16 chinese_name: "通义千问Qwen2.5-14B-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/Qwen2.5-14B-Instruct base_model_relation: finetune parameters: 4.9B tensor_type: - BF16 - I32 - 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: 258 stars: 1 tags: - Chat - Instruct --- # Qwen2.5-14B-Instruct-Int8-W8A16 > Qwen2.5-14B-Instruct-Int8-W8A16 - okwinds 在 ModelScope 开源的模型。通义千问2.5-14B-Instruct-INT8-W8A16 okwinds/Qwen2.5-14B-Instruct-Int8-W8A16 是 ModelScope 魔搭社区上的 4.9B 参数text-generation模型，采用 qwen 许可，基于 Qwen/Qwen2.5-14B-Instruct 构建，可用 deploy_task text/emb、lmdeploy 0.9.1、lmdeploy_turbomind 0.9.1 部署。 - **Repository**: okwinds/Qwen2.5-14B-Instruct-Int8-W8A16 - **License**: qwen - **Tasks**: text-generation - **Parameters**: 4.9B - **Base model**: Qwen/Qwen2.5-14B-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, Instruct - **Downloads**: 258 - **Stars**: 1 - **Last updated**: 2025-04-17 Source: https://www.modelscope.cn/models/okwinds/Qwen2.5-14B-Instruct-Int8-W8A16 --- # 通义千问2.5-14B-Instruct-INT8-W8A16 原模型 [qwen/Qwen2.5-14B-Instruct](https://www.modelscope.cn/models/qwen/Qwen2.5-14B-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/Qwen2.5-14B-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2.5-14B-Instruct-INT8-W8A16 是一个基于 通义千问2.5-14B-Instruct 的 INT8 量化并校准的模型。 - **模型名称:** Qwen2.5-14B-Instruct-Int8-W8A16 - **模型架构:** Qwen2.5 - **权重量化:** INT8 该模型通过将 [Qwen2.5-14B-Instruct](https://www.modelscope.cn/models/qwen/Qwen2.5-14B-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 >>> vllm serve "/home/gavin/llm/Qwen2.5-14B-Instruct-Int8-W8A16" --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.9 --served-model-name "Qwen2.5-14B-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 | Qwen2.5-14B-Instruct | Int8-W8A16-v1.0.0 | Recovery-v1.0.0 | _Int8-W8A16-master(New)_ | Recovery-master(New) | | ------------------- | -------------------- | ----------------- | --------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------- | |**Overall**|72.59|72.64|100.07%|72.67|100.12%| |Exam|82.67|82.57|99.88%|82.73|100.08%| |Language|57.40|57.14|99.54%|57.43|100.05%| |Knowledge|59.94|60.52|100.97%|59.95|100.02%| |Understanding|72.66|72.50|99.78%|72.71|100.07%| |Coding|80.57|80.16|99.48%|80.07|99.38%| |Reasoning|71.88|71.70|99.75%|71.86|99.97%| |Instruct Follow|82.99|83.89|101.08%|83.95|101.16%| |**Exam**|-|-|-|-|-| |C-Eval|81.38|81.05|99.59%|81.61|100.28%| |MMLU|79.44|79.62|100.23%|79.52|100.10%| |CMMLU|82.75|82.80|100.06%|82.85|100.12%| |GaokaoBench|87.09|86.79|99.66%|86.93|99.82%| |**Language**|-|-|-|-|-| |WiC|56.74|56.74|100.00%|56.74|100.00%| |WSC|81.73|80.77|98.83%|81.73|100.00%| |Afqmc-DEV|70.71|70.67|99.94%|70.69|99.97%| |Tydiqa-Goldp|20.42|20.36|99.71%|20.57|100.73%| |**Knowledge**|-|-|-|-|-| |BoolQ|84.04|84.10|100.07%|84.13|100.11%| |TriviaQA|56.89|56.56|99.42%|56.82|99.88%| |GPQA Diamond|38.89|40.91|105.19%|38.89|100.00%| |**Understanding**|-|-|-|-|-| |C3|93.92|93.75|99.82%|93.92|100.00%| |Race-Middle|93.11|92.90|99.77%|93.38|100.29%| |Race-High|89.79|89.57|99.75%|89.71|99.91%| |LCSTS|13.82|13.77|99.64%|13.82|100.00%| |**Coding**|-|-|-|-|-| |Human Eval|82.93|81.71|98.53%|82.32|99.26%| |Sanitized MBPP|78.21|78.60|100.50%|77.82|99.50%| |**Reasoning**|-|-|-|-|-| |Ocnli|68.31|67.86|99.34%|67.80|99.25%| |COPA|98.00|97.00|98.98%|98.00|100.00%| |MATH|51.82|51.00|98.42%|51.34|99.07%| |GSM8K|89.84|89.61|99.74%|90.52|100.76%| |Winogrande|78.85|79.32|100.60%|78.77|99.90%| |Hellaswag|90.71|90.60|99.88%|90.60|99.88%| |BBH|75.88|76.21|100.43%|76.58|100.92%| |TheoremQA|21.62|22.00|101.76%|21.25|98.29%| |**Instruct Follow**|-|-|-|-|-| |IFEval|82.99|83.89|101.08%|83.95|101.16%| _以上 benchmark 在 Nvidia L40 上完成，量化工作在 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 0.6.1.post2 - 原模型 BF16 峰值并发，在无 pending 的前提下，达到 running 100左右。(*Nvidia L40*) - 量化后的本模型 INT8（W8A16） 峰值并发，在无 pending 的前提下，达到 running 130左右。 (*Nvidia 4090 RTX*)
_______________________________ #### _作者在此 👇🏻 扫一扫_