Do not fuse gelu into FP8 cublasLT matmul on CUDA versions less than 12.4. #64245
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Do not fuse gelu into FP8 cublasLT matmul on CUDA versions less than 12.4.
This causes cublas LT to give an error in certain cases. Specially, it always causes an error on Ada and causes an error when fast accumulation is disabled on Hopper, at least in the cases I've tried. I am just disabling it in general for older CUDA versions in case there are other scenarios where it fails.
The issue actually only occurs on CUDA versions less than 12.3.2, but there is no good way of checking the patch version of CUDA, so I'm disabling it on CUDA versions less than 12.4. There is a CUBLAS_VER_PATCH with the patch version of cublas but this is undocumented so I'd rather not use it.
Thank you @philipphack and @Artem-B for helping me debug this.