vulkan: use 4 rows for scalar FA large tile size#18033
Closed
jeffbolznv wants to merge 1 commit intoggml-org:masterfrom
Closed
vulkan: use 4 rows for scalar FA large tile size#18033jeffbolznv wants to merge 1 commit intoggml-org:masterfrom
jeffbolznv wants to merge 1 commit intoggml-org:masterfrom
Conversation
github-actions
bot
added
Vulkan
jeffbolznv
force-pushed
the
fa_scalar_num_large_rows_4
branch
from
874549c to
cdb09db
Compare
Contributor
|
Here are performance results from my hardware, sadly it's not this easy. This change causes not only a few improvements, but also some regressions: Nvidia RTX 3090
AMD Radeon 8060S
AMD Radeon Pro VII
I wanted to test Intel, too, but I have some kind of memory retention issue with the driver currently, it gives me OOM on extended tests. |
Contributor
Author
|
Hmm, I thought the testing had generally been positive at #17715 (comment). But maybe it won't be so easy. |
Contributor
|
Yes, but I only checked 0 depth, I should have also run tests with context. How did you run it? |
Contributor
Author
|
Pretty sure I just tested with 0 depth. |
Contributor
|
Maybe we can pick the 8-row shader for larger context. I'll look into the difference in FA calls, maybe there's a better way to select the shader. |
Contributor
|
Here's a perf logger comparison for gpt-oss 20B at 8192 context: AMD 8060S (without coopmat)
AMD Radeon Pro VII
Nvidia RTX 3090
Looks like ubatch (ne2) and cache size (nek1) are most relevant to whether 4 or 8 is better. |
Contributor
Author
|
Using fewer rows is generally going to be worse for large batches (pp) because the KV cache has to be redundantly fetched more times. But it can be better for small batches (tg) when the group query attention factor is small, or if the larger number of rows doesn't fit in register file (the original case I think we were hitting on Intel). But it's clear that this change as-is is too simple. |
This was referenced Dec 22, 2025
Contributor
|
Superceded by #18280 |
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
2 participants
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.
See #17715 (comment). I also tested locally on 5090 in scalar mode and it was a few percent faster on several models.