[SYCL] Enabled more data types for oneMKL's gemm_batch API #8236
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
github-actions
bot
added
ggml
AidanBeltonS
approved these changes
Jul 1, 2024
|
oneapi-src/oneMKL#466 is merged last week, shall you wait for the the next release of oneMKL? Sorry I am not familiar with oneMKL. |
|
@airMeng Thanks for the suggestion. At the moment there is no clear/official release process on oneMKL Interface side. We don't mention anything related to oneMKL Interface releases in the README-sycl.md as well so from a user perspective it shouldn't be confusing at the moment. |
mofosyne
added
medium severity
airMeng
approved these changes
Jul 5, 2024
4 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Additional
gemm_batchtypes have been enabled in oneMKL (oneapi-src/oneMKL#466) and this patch enables their corresponding APIs for the SYCL backend which eliminates the extra-steps needed when targetting NON INTEL devices to cast/copy input/output to the supported types.The enablement of
gemm_batch_impl<sycl::half, sycl::half, float, float>for instance removes the overhead of callinggemm_batch_impl<sycl::half, sycl::half, sycl::half, sycl::half>followed by ato_fp32_syclfor thedstto be copied back fromfp16tofp32, which directly affects the KQ + KQV multi-batch path in quantized models Prompt Processing for instance.Performance on intel GPUs remain the same, and a slight improvement in Prompt Processing performance on some Nvidia GPUs was observed (0 to 3% on average).