Skip to content
This repository has been archived by the owner on Mar 21, 2024. It is now read-only.

Need a macro to disable bf16 support? #478

Closed
leofang opened this issue May 11, 2022 · 2 comments · Fixed by #480
Closed

Need a macro to disable bf16 support? #478

leofang opened this issue May 11, 2022 · 2 comments · Fixed by #480
Assignees
Labels
nvbug Has an associated internal NVIDIA NVBug. P1: should have Necessary, but not critical. type: enhancement New feature or request.
Milestone

Comments

@leofang
Copy link
Member

leofang commented May 11, 2022

After #306 CUB supports bf16 in some functions. Unfortunately, this means the bf16 headers (cuda_bf16.h, cuda_bf16.hpp) become a dependency of Thrust/CUB. Let me explain why it is unfortunate.

Normally, this would be fine in the majority of cases, as CUDA headers are also needed at compile time. But in some cases, such as within a CUDA runtime docker (which contains no headers) + using Jitify to compile CUB kernels at runtime, it is a big issue due to the missing headers.

Usually, the community approach is to bundle certain headers and redistribute them, so for example Thrust, CUB, and the fp16 headers (cuda_fp16.h, cuda_fp16.hpp) can be (and are being) repackaged in many projects. The unfortunate point is that the bf16 headers are not redistributable according to the CUDA EULA, and based on an internal conversation it'd remain the case for a long while. Therefore, it would be nice if CUB could offer a macro to disable the bf16 support.

@gevtushenko
Copy link
Collaborator

@leofang thank you for reporting the issue! Please, check if the following PR addresses your issue.

@leofang
Copy link
Member Author

leofang commented May 11, 2022

Relevant: nvbugs 3641496

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
nvbug Has an associated internal NVIDIA NVBug. P1: should have Necessary, but not critical. type: enhancement New feature or request.
Projects
None yet
Development

Successfully merging a pull request may close this issue.

3 participants