Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
测试QBMM int4图模式时，报错output type应为fp16/bf16但实际为fp32
Ascend950、cann-9.0.0
构建并运行一个输出类型为fp16的QBMM节点
编译正常，精度正常
When input is INT4/INT8 with pertokenScale, output dtype should be FLOAT16/BFLOAT16 actual is DT_FLOAT.
/assign
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Describe the current behavior / 问题描述 (Mandatory / 必填)
测试QBMM int4图模式时，报错output type应为fp16/bf16但实际为fp32
Environment / 环境信息 (Mandatory / 必填)
Ascend950、cann-9.0.0
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
构建并运行一个输出类型为fp16的QBMM节点
Describe the expected behavior / 预期结果 (Mandatory / 必填)
编译正常，精度正常
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
When input is INT4/INT8 with pertokenScale, output dtype should be FLOAT16/BFLOAT16 actual is DT_FLOAT.
Special notes for this issue/备注 (Optional / 选填)