[Bug]: Qwen3-30B-A3B模型执行w4a8量化叠加gptq量化算法，量化报错ERROR - Original exception: ZeroDivisionError('division by zero')-MindStudio-ModelSlim-AtomGit

已关闭

[Bug]: Qwen3-30B-A3B模型执行w4a8量化叠加gptq量化算法，量化报错ERROR - Original exception: ZeroDivisionError('division by zero') #30

mominhua创建于 2月6日关闭于 2月13日

mominhua成员

2月6日创建

在提交新问题之前，请确保您已经在社区中搜索过相关问题，并使用了社区中提供的资源/工具后，仍未找到满意的解决方式。

⚠️ 安全信息提醒：请仔细检查提供的文本内容，确保其不包含敏感数据信息，包括但不限于：

API 令牌或密钥
密码或身份验证凭证
私有网址或接口地址
个人或机密数据
...

在分享配置信息或代码示例时，请将敏感信息脱敏处理，或使用 <TOKEN> 等占位符替代原有内容。

环境信息

例如：
- 操作系统
- 昇腾硬件信息
- CANN软件版本
- 安装的对应软件版本

msmodelslim拉取的本代码仓master源码
环境为A2环境

🐛 问题描述

yaml配置如下：
apiversion: modelslim_v1

default_w8a8_dynamic: &default_w8a8_dynamic
act:
scope: "per_token"
dtype: "int8"
symmetric: True
method: "minmax"
weight:
scope: "per_channel"
dtype: "int8"
symmetric: True
method: "ssz"

default_w4a8_dynamic: &default_w4a8_dynamic
act:
scope: "per_token"
dtype: "int8"
symmetric: True
method: "minmax"
weight:
scope: "per_channel"
dtype: "int4"
symmetric: True
method: "gptq"
ext:
group_size: 64

spec:
process:
- type: "flex_smooth_quant"
enable_subgraph_type:
- 'norm-linear'
- 'ov'
include:
- ""
- type: "group"
configs:
- type: "linear_quant"
qconfig: default_w8a8_dynamic
include: ["self_attn"]
- type: "linear_quant"
qconfig: default_w8a8_dynamic
include: ["mlp"]
exclude: ["gate", "mlp.experts."]
- type: "linear_quant"
qconfig: default_w8a8_dynamic
include: ["model.layers.41.mlp.experts", "model.layers.41.mlp.experts", "model.layers.42.mlp.experts", "model.layers.43.mlp.experts", "model.layers.44.mlp.experts", "model.layers.45.mlp.experts", "model.layers.46.mlp.experts", "model.layers.47.mlp.experts"]
- type: "linear_quant"
qconfig: default_w4a8_dynamic
include: ["mlp.experts"]
exclude: ["model.layers.41.", "model.layers.42.", "model.layers.43.", "model.layers.44.", "model.layers.45.", "model.layers.46.", "model.layers.47."]

save:
- type: "ascendv1_saver"
part_file_size: 4

执行量化命令：
msmodelslim quant --model_path /packages/models_weight/Qwen3-30B-A3B/ --save_path /home/msmodelslim/model_quant/Qwen3-30B-A3B-w4a8-gptq-per-token-per-channel --model_type Qwen3-30B --device npu --config_path /home/msmodelslim/yaml/qwen3-30b-w4a8-per_token_gptq_per_channel.yaml --trust_remote_code True

量化报错：
2026-02-06 15:07:52,674 - msmodelslim.app.naive_quantization - ERROR - [UnexpectedError] Code: 500, Message: Unexpected error., TIP: Please report this issue to the msModelSlim developers. Repository: https://gitcode.com/Ascend/msmodelslim Q&A: https://gitcode.com/Ascend/msmodelslim/blob/master/docs/zh/faq.md
2026-02-06 15:07:52,674 - msmodelslim.app.naive_quantization - ERROR - Original exception: ZeroDivisionError('division by zero')