Mirror of Eco-Tech/DeepSeek-V4-Flash-w8a8-mtp from ModelScope. All credit to the original authors. Not an official release from them. License unchanged (mit).

DeepSeek-V4-Flash-w8a8-mtp

1. 基本信息

项目 信息
原始模型名 DeepSeek-V4-Flash
原始模型链接 deepseek-ai/DeepSeek-V4-Flash
精度测试机型 Atlas 800T A2 1台
精度测试平台 docker vllm-ascend
版本 vllm-ascend:v0.13.0rc3
链接 quay.m.daocloud.io/ascend/vllm-ascend:v0.13.0rc3

2 量化脚本:

现已集成一键量化

msmodelslim quant \
 --model_path ${model_path} \
 --save_path ${save_path} \
 --model_type DeepSeek-V4-Flash \
 --quant_type w8a8 \
 --trust_remote_code True

3 精度测试结果

模型名 量化格式 数据集 测试精度 % 官方精度 % 备注
DeepSeek-V4-Flash-w8a8-mtp w8a8 gpqa 71.21 71.2 V4-Flash Non-Think
DeepSeek-V4-Flash-w8a8-mtp w8a8 mmlupro 82.85 83.0 V4-Flash Non-Think
DeepSeek-V4-Flash-w8a8-mtp w8a8 mmlupro 85.86 86.2 V4-Flash Max
  • 使用ais_bench,其中Non-Think模式max_out_len = 65536, Max模式max_out_len = 131072。精度存在波动,建议多次测试。

4 思考模式开启方法

4.1 Curl指令:

Non-Think: 不加思考参数

High: "chat_template_kwargs": {"thinking": true, "reasoning_effort": "high"}

Max: "chat_template_kwargs": {"thinking": true, "reasoning_effort": "max"}

4.2 Ais_bench Benchmark:

Non-Think: 不加思考参数

High:

generation_kwargs=dict(
            ....
            chat_template_kwargs = {"thinking": True, "reasoning_effort": "high"}
        )

Max:

generation_kwargs=dict(
            ....
            chat_template_kwargs = {"thinking": True, "reasoning_effort": "max"}
        )
Downloads last month
414
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ml-resources/DeepSeek-V4-Flash-w8a8-mtp

Quantized
(135)
this model

Collection including ml-resources/DeepSeek-V4-Flash-w8a8-mtp