HIP: Enable fattn-mma kernel on cdna for dkq > 256 for large batch sizes - #28907
Merged
Merged
Conversation
JohannesGaessler
approved these changes
Sep 15, 2026
JohannesGaessler
left a comment
Contributor
There was a problem hiding this comment.
| GPU | Model | Microbatch size | Test | t/s b10952 | t/s a7899f0 | Speedup |
|---|---|---|---|---|---|---|
| MI100 | deepseek2 16B Q4_0 | 1 | pp1024@d32768 | 88.53 | 89.16 | 1.01 |
| MI100 | deepseek2 16B Q4_0 | 2 | pp1024@d32768 | 95.42 | 95.35 | 1.00 |
| MI100 | deepseek2 16B Q4_0 | 4 | pp1024@d32768 | 126.17 | 126.23 | 1.00 |
| MI100 | deepseek2 16B Q4_0 | 8 | pp1024@d32768 | 195.22 | 194.84 | 1.00 |
| MI100 | deepseek2 16B Q4_0 | 16 | pp1024@d32768 | 257.18 | 312.10 | 1.21 |
| MI100 | deepseek2 16B Q4_0 | 32 | pp1024@d32768 | 314.76 | 407.08 | 1.29 |
| MI100 | deepseek2 16B Q4_0 | 64 | pp1024@d32768 | 332.68 | 453.24 | 1.36 |
| MI100 | deepseek2 16B Q4_0 | 128 | pp1024@d32768 | 378.64 | 551.63 | 1.46 |
| MI100 | deepseek2 16B Q4_0 | 256 | pp1024@d32768 | 412.25 | 580.00 | 1.41 |
| MI100 | deepseek2 16B Q4_0 | 512 | pp1024@d32768 | 435.17 | 731.22 | 1.68 |
IMbackK
force-pushed
the
cdnammafattn256
branch
from
September 15, 2026 20:09
fd9e3b1 to
d21d8d0
Compare
Contributor
Author
|
force push was for: braces where unbalanced after applying suggestion above. |
IMbackK
force-pushed
the
cdnammafattn256
branch
from
September 21, 2026 15:04
d21d8d0 to
458b113
Compare
Contributor
|
Is it intentional that you removed the merge ready label again? |
Contributor
Author
|
it was, since the quality check was failing on vgpr spills, but since i added them to the ignore list this can be merged now. |
pierreguillot
pushed a commit
to Ircam-Partiels/llama.cpp
that referenced
this pull request
Oct 1, 2026
…zes (ggml-org#28907) * HIP: Enable fattn-mma kernel on cdna for dkq > 256 for large batch sizes * CI: hip-quality-check: ignore spills for very large mfma mma kernels
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
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.
Overview
For large head sizes the mma kernel still eventually outscales the tile kernel even at occupancy 1 limited by smem
Additional information
Details
<\details>
Requirements