Gather qmm batched kernel and refactoring of quantized #2078
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
jagrit06
reviewed
Apr 16, 2025
barronalex
approved these changes
Apr 16, 2025
| int N = out.shape(-1); | ||
|
|
||
| int vector_limit = transpose_ ? get_qmv_batch_limit(K, N, d) : 4; | ||
|
|
There was a problem hiding this comment.
This is so much easier to read. Nice job!
|
For DeepSeek with its 256 experts we have a smaller speedup for small prompts and we 're back at 2x when there are enough tokens to fill the experts.
As expected I see no regression in non MoE models so I think this is good to merge. |
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Outdated suggestions cannot be applied.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.
This is part 2 of the #2040 kind of. It implements the same kernel but for quantized mm and refactors
quantized.cpp. It also sets upquantized.hfor a refactor of the batched kernel (mergeqmm_nandqmm_tand more).In the process of refactoring it fixes the following 2 bugs
gather_qmvtoqmv_quadqvmIn the microbenchmark we see similar speedups as for #2040. A little less impressive because
qmvis faster and we have to skip less data to move from matrix to matrix. I am adding some real world MoE prompt processing speedups below