ibsidorenko opened a new pull request, #16892: URL: https://github.com/apache/tvm/pull/16892
This is attempt to bring [commit](https://github.com/octoml/tvm/commit/f21b9c9c561e7bcb7a81ae12d71568c6e7c1fc49) and align `octoml/tvm` with `apache/tvm`. This commit replaces fp16 compute dtype and scale dtype by fp32 in cublas matmul. According to cuBLAS [docs](https://docs.nvidia.com/cuda/cublas/index.html#cublasltmatmul) there are two possible options for compute/scale dtype when input/output dtype is fp16: 1. compute dtype is `fp16` and scale dtype is `fp16` 2. compute dtype is `fp32` and scale dtype is `fp32` By default, we use 1) in apache/tvm and 2) in octoml/tvm. This commit aligns different behaviour and set `fp32` as default. cc @vinx13 @masahi -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
