HIP: enable WMMA-MMQ INT kernels for RDNA 3#17576
Merged
JohannesGaessler merged 5 commits intoggml-org:masterfrom Dec 5, 2025
Merged
Conversation
ggml/src/ggml-cuda/mma.cuh
Outdated
| static constexpr int ne = I * J / 32; | ||
| #elif defined(RDNA3) | ||
| static constexpr int ne = (I == 16 && J == 16) ? I * J / 32 : I * J / 16; | ||
| #endif |
Contributor
There was a problem hiding this comment.
Suggested change
| #endif | |
| #endif // defined(RDNA4) |
Please add comments to indicate which #if/#ifdef and #endif is closing.
ggml/src/ggml-cuda/mmq.cu
Outdated
Comment on lines
310
to
312
| if (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) { | ||
| return true; | ||
| } |
Contributor
There was a problem hiding this comment.
Suggested change
| if (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc)) { | |
| return true; | |
| } | |
| return true; |
ggml/src/ggml-cuda/mmq.cuh
Outdated
Comment on lines
+1545
to
+1548
| A1.x[0] = 0x01010101; | ||
| A1.x[1] = 0x01010101; | ||
| A1.x[2] = 0x01010101; | ||
| A1.x[3] = 0x01010101; |
Contributor
There was a problem hiding this comment.
Suggested change
| A1.x[0] = 0x01010101; | |
| A1.x[1] = 0x01010101; | |
| A1.x[2] = 0x01010101; | |
| A1.x[3] = 0x01010101; | |
| #pragma unroll | |
| for (int l = 0; l < tile_A::ne; ++l) { | |
| A1.x[l] = 0x01010101; | |
| } |
To my understanding tile_A has 4 elements for RDNA3 but for RDNA4 it only has 2 elements. So as it is this would result in out-of-bounds writes and potential memory trampling for RDNA4.
Contributor
Performance
In terms of performance I think this PR would be good to merge. There are some cases around batch size 32 that have suboptimal performance but that particular batch size is comparatively less important vs. the larger ones. So I think it would be fine to merge the PR as-is and to maybe optimize that use case in a follow-up PR. (Batch sizes 1-8 are using the same code for both tests so changes there are just random noise and can be ignored, I only included them to investigate the scaling.) |
JohannesGaessler
approved these changes
Dec 3, 2025
Contributor
|
(This PR still needs a rebase on top of master.) |
a34b76f to
c9ec96c
Compare
Contributor
Author
has been done |
ggml/src/ggml-cuda/mma.cuh
Outdated
| ); | ||
| #endif // defined(RDNA4) | ||
|
|
||
| #elif defined(RDNA3) |
Contributor
There was a problem hiding this comment.
Suggested change
| #elif defined(RDNA3) | |
| #elif defined(RDNA3) |
To fix the EditorConfig CI.
5941226 to
685be0e
Compare
Member
Contributor
|
Seems like some minor issues on RDNA4: hjc4869@df264e1 |
Contributor
|
Thank you for this! Those are great speed up results. I did get some errors and my build failed, could you please take a look? |
Merged
Contributor
|
On GFX1100 + ROC641 seems like this commit is causing |
Contributor
Author
I believe this is due to FP16/BF16 MMF kernels not been enabled yet, once this pr (#17495) get merged it should no longer get this failure |
Contributor
I suppose ROCm builds will just be broken for RDNA 3 until someone finds the time to finish that PR then? |
This was referenced Dec 6, 2025
Beinsezii
added a commit
to Beinsezii/llama.cpp
that referenced
this pull request
Dec 6, 2025
This reverts commit 668ed76.
JayZenith
pushed a commit
to JayZenith/llama.cpp
that referenced
this pull request
Dec 7, 2025
* enabled wmma instructions for most quantizations other than q2k * fixed the last q2_k test case failure * address comments: fix out of bound write for RDNA4, add comments after #endif * clean up rebase: fix ne error in half2 * fix the EditorConfig CI
0Marble
pushed a commit
to 0Marble/llama.cpp
that referenced
this pull request
Dec 18, 2025
* enabled wmma instructions for most quantizations other than q2k * fixed the last q2_k test case failure * address comments: fix out of bound write for RDNA4, add comments after #endif * clean up rebase: fix ne error in half2 * fix the EditorConfig CI
Nexesenex
added a commit
to Nexesenex/croco.cpp
that referenced
this pull request
Dec 20, 2025
This reverts commit 668ed76.
Anico2
added a commit
to Anico2/llama.cpp
that referenced
this pull request
Jan 15, 2026
* enabled wmma instructions for most quantizations other than q2k * fixed the last q2_k test case failure * address comments: fix out of bound write for RDNA4, add comments after #endif * clean up rebase: fix ne error in half2 * fix the EditorConfig CI
blime4
referenced
this pull request
in blime4/llama.cpp
Feb 5, 2026
* enabled wmma instructions for most quantizations other than q2k * fixed the last q2_k test case failure * address comments: fix out of bound write for RDNA4, add comments after #endif * clean up rebase: fix ne error in half2 * fix the EditorConfig CI
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
6 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.
Enabled WMMA-MMQ INT kernels for RDNA 3 architecture on AMD GPUs
Following similar approach to #17156
Using ./build/bin/llama-bench to collect the following performance results
Build command for the following performance results:
HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" cmake -S . -B build -DGGML_HIP=ON -DGGML_CUDA_FORCE_MMQ=OFF -DGGML_HIP_UMA=OFF -DGGML_HIP_ROCWMMA_FATTN=OFF -DGPU_TARGETS="gfx1100" -DGGML_HIP_GRAPHS=OFF -DLLAMA_CURL=OFF -DGGML_CUDA_FORCE_CUBLAS=OFF -DCMAKE_BUILD_TYPE=Release && cmake --build build --config Release -- -j 32
Popular models performance result for AMD Radeon AI PRO W7900 (gfx1100)
Popular models performance result for AMD Strix Halo (gfx1151)
All quantization performance result for AMD Radeon AI PRO W7900 (gfx1100)
All quantization performance result for AMD Strix Halo (gfx1151)