towards an even better backward attention kernel #179
Merged
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
ngc92
force-pushed
the
even-better-attention
branch
from
5f1b7a3
to
51c6d8a
Compare
karpathy
reviewed
Apr 19, 2024
| @@ -560,6 +561,128 @@ __global__ void softmax_autoregressive_backward_kernel5(float* __restrict__ dpre | |||
| } | |||
| } | |||
|
|
|||
|
|
|||
| // I want `BlockSize` to be statically known to the compiler, thus we get a template here. | |||
ngc92
force-pushed
the
even-better-attention
branch
from
d8308d6
to
50e105a
Compare
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Outdated suggestions cannot be applied.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.
Back to the drawing board, because I think the other kernels hit a local minimum, or at least the way the loops where organized made it very difficult to think about how to optimize this further.
I think there is quite a bit of room to optimize this version further, but for educational purposes, I think it is better to have a simpler version first, where the main idea is evident, and then go crazy with the optimizations in a second version.
This design limits T to multiples of the block_size. Incidentally, on my system at least, block size 64 turns out to be fastest, which actually is compatible with both our training at test scripts.