First pass in simplifying and improving qmm #1030
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
jagrit06
reviewed
Apr 24, 2024
Comment on lines
647
to
655
|
|
||
| // Instantiate the appropriate BlockMMA and Loader | ||
| using mma_t = mlx::steel::BlockMMA<T, T, BM, BN, BK, WM, WN, false, true, BK, BK>; | ||
| using loader_x_t = mlx::steel::BlockLoader<T, BM, BK, BK, 1, WM * WN * SIMD_SIZE, 1, 4>; | ||
| using loader_x_t = mlx::steel::BlockLoader<T, BM, BK, BK, 1, WM * WN * SIMD_SIZE>; | ||
| using loader_w_t = QuantizedBlockLoader<T, BN, BK, BK, 1, WM * WN * SIMD_SIZE, group_size, bits>; | ||
|
|
||
| threadgroup T scales_block[BN * groups_per_block]; | ||
| threadgroup T biases_block[BN * groups_per_block]; | ||
| threadgroup T Xs[BM * BK]; | ||
| threadgroup T Ws[BN * BK]; | ||
|
|
There was a problem hiding this comment.
Changing the leading dim of the threadgroup memory should help with bank conflicts
Doing so should be as simple as
constexpr int BK_padded = (BK + 16 / sizeof(T));
// Instantiate the appropriate BlockMMA and Loader
using mma_t = mlx::steel::BlockMMA<T, T, BM, BN, BK, WM, WN, false, true, BK_padded, BK_padded>;
using loader_x_t = mlx::steel::BlockLoader<T, BM, BK, BK_padded, 1, WM * WN * SIMD_SIZE>;
using loader_w_t = QuantizedBlockLoader<T, BN, BK, BK_padded, 1, WM * WN * SIMD_SIZE, group_size, bits>;
threadgroup T Xs[BM * BK_padded];
threadgroup T Ws[BN * BK_padded];
awni
approved these changes
Apr 24, 2024
|
After @jagrit06 's improvement to avoid memory bank conflicts and remove the checks out of the loop we now get for a total improvement of 30%-40%! |
|
Incredible!! 🔥 |
|
WOW, just WOW |
1 task
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.
Extracted the loading and dequantizing logic to a reusable
QuantizedBlockLoaderà lasteel. Also removed the reading of scales/biases in smem and reduced the tile size that provides a pretty big speedup.Before on Mistral 7B 4bit 64 group size:
after
all in all a pretty consistent 20%-25% speedup.