CUDA: enable Gemma FA for HIP/Pascal #9581
Merged
Conversation
This file contains hidden or 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
testing
slaren
approved these changes
Sep 21, 2024
dsx1986
pushed a commit
to dsx1986/llama.cpp
that referenced
this pull request
Oct 29, 2024
arthw
pushed a commit
to arthw/llama.cpp
that referenced
this pull request
Nov 15, 2024
arthw
pushed a commit
to arthw/llama.cpp
that referenced
this pull request
Nov 18, 2024
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Outdated suggestions cannot be applied.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.
Fixes #9580 .
As of right now the CUDA backend reports that for FlashAttention a head size of 256 is only supported for NVIDIA GPUs that are Volta or newer. However, for AMD GPUs and old NVIDIA GPUs a head size of 256 can also be enabled if the vector kernel is used for large batch sizes. The performance won't be great but it will be faster than CPU. This PR adapts the CUDA code to enable this.
Also I noticed that the tests were only testing batch sizes < 8 which meant that some CUDA kernels were not being invoked at all. I changed the batch sizes to cover a wider range.