SYCL: Add mrope kernel #13755
Merged
SYCL: Add mrope kernel #13755
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
Alcpz
reviewed
May 29, 2025
ggml/src/ggml-sycl/rope.cpp
Outdated
Comment on lines
141
to
142
| dst[i + 0] = x[i + 0]; | ||
| dst[i + 1] = x[i + 1]; |
There was a problem hiding this comment.
Suggested change
| dst[i + 0] = x[i + 0]; | |
| dst[i + 1] = x[i + 1]; | |
| *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i); |
I've tried checked this change, and for big enough tensors it makes a noticeable difference in performance:
dst[i + 0] = x[i + 0];
dst[i + 1] = x[i + 1];
ROPE(type=f32,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 14944 runs - 68.72 us/run - 71995 kB/run - 1007.63 GB/s
ROPE(type=f32,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 4400 runs - 232.89 us/run - 167995 kB/run - 701.70 GB/s
ROPE(type=f32,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 14944 runs - 68.82 us/run - 71995 kB/run - 1006.27 GB/s
ROPE(type=f32,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 4400 runs - 232.11 us/run - 167995 kB/run - 704.05 GB/s
ROPE(type=f16,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 46650 runs - 21.58 us/run - 35997 kB/run - 1597.78 GB/s
ROPE(type=f16,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 11200 runs - 92.34 us/run - 83997 kB/run - 876.16 GB/s
ROPE(type=f16,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 47583 runs - 21.43 us/run - 35997 kB/run - 1608.58 GB/s
ROPE(type=f16,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 10800 runs - 92.64 us/run - 83997 kB/run - 873.38 GB/s
*reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
ROPE(type=f32,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 21015 runs - 47.71 us/run - 71995 kB/run - 1451.41 GB/s
ROPE(type=f32,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 6000 runs - 167.92 us/run - 167995 kB/run - 973.19 GB/s
ROPE(type=f32,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 21482 runs - 47.53 us/run - 71995 kB/run - 1456.84 GB/s
ROPE(type=f32,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 6000 runs - 168.06 us/run - 167995 kB/run - 972.38 GB/s
ROPE(type=f16,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 50382 runs - 20.08 us/run - 35997 kB/run - 1716.75 GB/s
ROPE(type=f16,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=0,v=1): 14800 runs - 68.47 us/run - 83997 kB/run - 1181.62 GB/s
ROPE(type=f16,ne_a=[1280,1200,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 49449 runs - 20.29 us/run - 35997 kB/run - 1699.40 GB/s
ROPE(type=f16,ne_a=[1280,2800,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.424500,ef=0.746500,af=1.424500,ff=1,v=1): 14800 runs - 68.92 us/run - 83997 kB/run - 1174.02 GB/s
I think it's worth considering changing it. Do you know which model makes use of this (Qwen2.5-VL?)? I'd be happy to grab the tk/s to see if this makes a significant change. I suppose it's not that impactful as our backend issues are mostly in the mul_mat
There was a problem hiding this comment.
From this comment Qwen 2.5 needs it.
There was a problem hiding this comment.
I have added this suggestion to other kernels as well. Thanks alot!
Uses `sycl::vec` to load and store two elements at a time, significantly improving performance in `rope_norm`, `rope_neox`, and `rope_multi`. This reduces the number of memory accesses and leverages SIMD instructions for faster execution.
Alcpz
approved these changes
May 29, 2025
8000
There was a problem hiding this comment.
I can't continue the review for a couple of days (in case you introduce other changes), but if @Rbiessy is happy with the changes, I'm ok with the merge. We can iterate over this in a different PR.
Edit: Not sure if someone else wants to review the code.
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 crea
2DE5
te 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.
No description provided.