[Issue]: ck_tile::type_convert<float> not working for half to float

[michael604work]

Problem Description

Hi,

While I was building with AITER, using MI350 & ROCM 7.02,
Faced some issues with

float a_fp32 = ck_tile::type_convert(a);
float b_fp32 = ck_tile::type_convert(b);
return ck_tile::type_convert(a_fp32 + b_fp32);

These, at csrc/include/custom_all_reduce.cuh.

Had to bypass those by changing to __half2float & float2half instead.

Operating System

CentOS Stream 9

CPU

AMD EPYC 9655

GPU

MI350X

ROCm Version

Rocm 7.0.2

ROCm Component

No response

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[TennyWang1223]

@michael604work
Hi, I ran the unit test for aiter custom_all_reduce in the MI350 & ROCm 7.02 environment. The fp16 dtype precision test passed without any issues. Could you please provide more detailed issue descriptions or error logs so we can understand what exactly happened? Also, please share detailed information about the runtime environment, such as the docker image name.