[ROCm] Add gfx1102/gfx1103 support - #40037
Merged
Merged
Conversation
… macros gfx1102 and gfx1103 (RDNA 3, e.g. Radeon 780M iGPU) were missing from the HIP_SUPPORTED_ARCHS list in CMakeLists.txt and from the compile-time architecture macros in skinny_gemms.cu and attention.cu. The skinny GEMM and attention kernels defined __HIP__GFX1X__ and __HIP__GFX11__ by listing individual arch targets (__gfx1100__, __gfx1101__, __gfx1150__, __gfx1151__) but missed gfx1102/gfx1103. This caused wvSplitK kernel bodies to compile as UNREACHABLE_CODE (assert false) while the wave32 reduction code was still active, resulting in a crash at runtime. Replace __HIP__GFX11__ / __HIP__GFX12__ with the compiler-provided __GFX11__ / __GFX12__ umbrella macros which are defined for every target in each family. This avoids needing to enumerate individual targets and automatically covers future gfx11xx/gfx12xx additions. Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
mgehre-amd
requested review from
LucasWilkinson,
gshtras,
tjtanaa and
tlrmchlsmth
as code owners
April 16, 2026 15:43
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request adds support for the gfx1102 and gfx1103 AMD GPU architectures in the CMakeLists.txt. It also refactors the ROCm source files to use standard compiler-defined macros, such as GFX11 and GFX12, replacing the previous custom macro definitions. I have no feedback to provide.
gshtras
approved these changes
Apr 17, 2026
avinashsingh77
pushed a commit
to avinashsingh77/vllm
that referenced
this pull request
Apr 27, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com> Signed-off-by: Avinash Singh <avinashsingh.rcoem@gmail.com>
Copilot AI
pushed a commit
to hongbolv/vllm
that referenced
this pull request
May 7, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com> Co-authored-by: hongbolv <33214277+hongbolv@users.noreply.github.com>
weifang231
pushed a commit
to weifang231/eb-vllm
that referenced
this pull request
May 13, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
my-other-github-account
pushed a commit
to my-other-github-account/vllm
that referenced
this pull request
May 15, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
mfylcek
pushed a commit
to mfylcek/vllm
that referenced
this pull request
May 19, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
brian-dellabetta
pushed a commit
to neuralmagic/vllm
that referenced
this pull request
May 29, 2026
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
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
3 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.
gfx1103 (RDNA 3, e.g. Radeon 780M iGPU) was missing from the HIP_SUPPORTED_ARCHS list in CMakeLists.txt and from the compile-time architecture macros in skinny_gemms.cu and attention.cu.
The skinny GEMM and attention kernels defined HIP__GFX1X and HIP__GFX11 by listing individual arch targets (gfx1100, gfx1101, gfx1150, gfx1151) but missed gfx1103. This caused wvSplitK kernel bodies to compile as UNREACHABLE_CODE (assert false) on gfx1103 while the wave32 reduction code was still active, resulting in a crash at runtime.
Replace HIP__GFX11 / HIP__GFX12 with the compiler-provided GFX11 / GFX12 umbrella macros which are defined for every target in each family. This avoids needing to enumerate individual targets and automatically covers future gfx11xx/gfx12xx additions.