Error in promotion of data type for accumulation
Open
Nobody has claimed this yet.
iris
- Dominant language
- Python
- Stars
- 202
- Forks
- 47
- Avg merge
- 6d 11h
- Merged PRs (30d)
- 4
Description
This piece of code in the iris/x/ kernels leads to error for many datatypes. Eg: fp16 get's accumulated into int32 leading to large RMSE errors in the output.
# Determine accumulator dtype based on output type
acc_dtype = tl.int32 if C.type.element_ty != tl.int8 else tl.float32
Eg. used here:
https://github.com/ROCm/iris/blob/1a09fae0572c2e2484f5abb0c09214a02ed4500b/iris/x/all_gather_gemm.py#L119C1-L119C73
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.